pr_review.py

#!/usr/bin/env python3
"""Consolidated Copilot-review status for a PR - one command, one compact digest.

Why this exists: 4,513 of the measured `gh` invocations were one call per agent turn, and
each turn re-bills the whole session context. The bytes `gh` returns are trivial, averaging
574, so the round-trips are the entire cost. This collapses a poll cycle into a single
invocation whose output is a few hundred bytes. See GOVERNANCE.md "Context and Delegation
Discipline" for the rule this implements.

Subcommands
  attest   Publish that a recorded local pass covers the pull request's head, as a comment carrying
           `<!-- fleet-local-review: head=<sha> base=<branch> findings=<n> -->` that `status` and
           `wait` read, an attestation counting only against the base it names. It reads the receipt
           `local_review.py` keeps in the checkout's git directory, which nothing on GitHub can, and
           refuses unless the checkout (--checkout, default the current directory) is at the pull
           request's head with no change beyond it, its merge base with the base branch is the pull
           request's own, and `local_review.py status` reports a covering pass against the base,
           read once for the coverage and the findings count alike. Exit 0 = posted, 64 = the write
           scope could not be established or excludes the target, 65 = the pull request could not be
           read, 66 = the response did not confirm the comment, 67 = the checkout is not the head,
           holds changes, or measures another merge base, or that merge base could not be read, 68 =
           no current local pass covers the content.
  comment  Post one PR-conversation answer, including a suppressed-finding disposition. The PR
           node id is read in the same run, and the returned comment URL and body confirm the
           write. Exit 0 = done, 64 = write scope could not be established or excludes the
           target, 65 = the PR could not be read, 66 = the response did not confirm the comment.
  claims   Check the description against the branch it describes. A body claiming a commit or
           quoting a `uses:` ref the head tree no longer carries is a silent failure caught by a
           reviewer or not at all, and three stale descriptions in one session generated six
           review findings between them. Read-only. Exit 0 = every reference resolves, 70 = one
           does not, 71 = there were references and none could be read, so nothing was decided.
  status   One digest line, any unresolved threads, and any suppressed findings. Read-only.
           Exit 0 = no Copilot review covers the head yet, or every output shape is recognized
           and full diff coverage is stated for this head, by the round covering it or by the
           carry below, or where nothing states or carries any, a round covering the head names
           exactly the changed files in its own file table, `coverage=table` in the digest, or
           where no round covering the head carries a table, the newest round that does names
           them under the same change-set bound, `coverage=carried:table`.
           `review_on_head` and `rounds=`
           in the digest name Copilot's own coverage specifically, the reviewer this script
           requests and waits for, never "no review of any kind covers this head": a
           tracked other reviewer, named under `other_reviewed` below, can carry the exact head
           commit while Copilot's own `review_on_head` still reads `NO`, and an empty body from
           that other reviewer on that head is its own ordinary "reviewed, nothing to flag"
           shape, the same reading an empty-bodied Copilot round already gets, not a gap.
           `review_on_head=local` reads a head no Copilot round covers on a pull request into a
           branch other than the default, after Copilot's first round, where an attestation vouches
           for it and no round on record states or appears to state partial coverage, with
           `coverage=local` beside it.
           Use `wait` when review presence is the condition, since `status` reports an absent
           review without treating it as a failure.
           42 = a round read fewer files than the pull request changed, so part of the diff
           has no review at all. That round is the one covering the head, or the one carried
           to it under 45 below, an unread part staying unread across a push that leaves the
           change set alone. Measured over four pull requests and seven rounds here, a
           re-request never cleared one and no round ever recovered, so this is a state to
           hand to the maintainer rather than one to retry into.
           43 = the reviewer sent a shape this script has no reader for, so no field here can
           be believed. The remedy is an issue on the repository hosting this script, and the
           review loop does not close until the reader is fixed. Merging regardless is the
           maintainer's decision rather than the agent's.
           45 = no statement of changed-file coverage describes this head, and no file table
           stands in for one. A round that
           states none carries the newest round that states some forward, bounded on the change
           set, so this covers three states: nothing ever stated coverage, the round that did
           describes a different set of changed files, or that comparison could not be read.
           Past those, a round covering the head whose own table names exactly the changed
           files stands in, or where none carries a table, the newest round that does, its table
           naming exactly the changed files, under the same change-set bound. Either table
           stands in only where no Copilot round on the pull request, on any commit, states or
           appears to state partial coverage, quoted or not, and the whole review
           history is in view, since over the rounds measured the table names the whole changed
           set on partial rounds too. The digest names why no table stood in.
           Copilot's Balanced review effort, the default since 2026-09-28, writes the table and
           almost never a statement, which is the case the table reading exists for.
           Where the head branch lacks the current review instructions, which a round at Lite
           effort needs to state coverage, bring them onto it, push, and run `wait`, which requests a round on the new head, and where the
           comparison could not be read, run `status` again. Past those, a develop -> main
           promotion can still satisfy coverage, in the one case pr-review-conduct's Merge Gate
           item 2 states. Otherwise hand the state to the maintainer rather than retrying into
           it, since a round
           re-requested on the same head states coverage or carries a table only by chance.
           50 = the round covering the head opens on a verdict other than the clean one, while it
           states no open finding, opened no thread, collapsed no suppressed or previously-missed
           finding, and no reviewer's thread is open, so what it flags is named in its headline
           alone, printed under `VERDICT WITH NO COUNTED FINDING`. Ranked under 42, 43, and 45,
           and over `wait`'s 44.
           The headline is owed the triage a suppressed finding gets, and the code repeats on that
           head after it is answered, since nothing here reads an answer.
           A refusal naming the account quota still reads as absent here, exit 0, since a refusal
           covers no head either. Its printed digest line carries `refusal=QUOTA` regardless, and
           `refusal=ERROR` for an error refusal whose run log names no rate limit or could not be
           read. Either field reads the pull request's newest Copilot review where it is a quota or
           error refusal, on an earlier head or over a head a genuine round covered before it, and a
           file-count refusal is spent by coverage of the same head. `wait` is where that state gets
           its own exit codes, 46 and 47 below, because only `wait` is the command a caller might
           otherwise poll out a timeout on.
           `unresolved` counts every open thread whoever opened it, since an open thread blocks
           a ruleset-gated merge whatever its author, and a filter on known reviewer logins
           once read `unresolved=0` over a qodo thread posted under a login it did not know.
           The known logins, Copilot, CodeRabbit (`coderabbitai`), and qodo
           (`qodo-free-for-open-source-projects`), only attribute a thread in the breakdown
           beside the count, where any other author reads as `other`.
           Both `threads=` and `unresolved=` are read from a single 100-thread page with no
           further pagination, so a pull request carrying more than that undercounts silently
           past that point: both fields print a trailing `+` and a `THREADS TRUNCATED` block
           follows, naming the gap rather than leaving either count to be trusted as whole.
           Reading past the cut needs `reply`'s own paginated walk instead. `suppressed=`
           and `cr_outside_diff=` carry the identical marker over the separate 100-review page
           they are both read from, a `REVIEWS TRUNCATED` block following, since an unanswered
           finding on a round old enough to fall out of that window is exactly as invisible.
           Coverage, refusal, and `wait`/request support stay Copilot-only, each being its own
           format and its own future task per bot. Where either has posted anything at all on
           the current head, `other_reviewed` names it, identity and commit only, no verdict
           read from what it said. `other_rate_limited` names one whose newest comment or review
           carries its own rate-limit marker instead, a structural
           `<!-- ... rate limited by ... -->` convention rather than free-text prose (observed on
           CodeRabbit), so reading it needs no per-bot wording model the way
           Copilot's quota refusal needed `QUOTA` built for its own text.
           Two of their own finding shapes are read, though, each its own blind spot a thread
           poll alone cannot see. `cr_outside_diff=N (on_head=X earlier=Y)` counts
           CodeRabbit's own "outside diff range" findings, collapsed into the review body rather
           than raised as an inline review comment because the finding sits on a line outside
           the pull request's changed hunks. Printed once CodeRabbit has raised one, on any
           round, or once the reviews window is truncated, since an older round could then
           still carry one unseen (`cr_outside_diff=0+`). Stays silent otherwise, whether that
           silence means no finding was raised or CodeRabbit was never trialed at all.
           `qodo_open=N` counts Qodo's own
           numbered findings that carry neither its `Resolved` nor `Dismissed` self-tracked
           badge: Qodo's formal review carries an empty body on every round observed, so its
           findings are read from its "Code Review by Qodo" PR-level comment instead, not
           head-scoped since a comment carries no commit. Printed as `0` once Qodo has posted
           that comment at all, since a `0` is itself a reading, `unknown` where its findings
           comment specifically could be sitting behind the 100-comment window even while a
           paired `PR Summary by Qodo` stays visible, so its silence cannot be told from it
           never having commented, and absent only where it genuinely never has. Qodo's own
           badge is a fast pre-triage signal, not a substitute for reading the finding:
           spot-verify against `gh pr diff` rather than trusting it outright.
           `overview=T/M` reads the second Copilot review-body format's own finding total, `T`,
           beside the number of review threads that round actually opened, `M`. `T` reads `?`
           where no total is found in the overview preamble or in a count-first open-findings
           section, which covers a round stating none, a round stating one only after its first
           collapsed section anywhere else, and a round stating one as a bullet, that last
           blocking at exit 43 as an unvetted metadata label. None of them says
           the round withheld nothing, and a `?` is a reason to read the body. Present
           only where the round covering the head is written in that format, which reached this
           repository after every round the vetted marker lists below were measured over.
           `M` is read from the API rather than from the review's prose. That format enumerates
           the findings it opened threads for, and the anchor each entry links is the database id
           of that thread's own comment, so the threads are the same set already structured.
           Reading the enumeration instead took six attempts and four of them cancelled a
           shortfall by counting something that was not an entry, telling those apart needing a
           Markdown parser rather than a line scan.
           A total larger than the thread count is findings the round raised that polling threads
           cannot see, and a `FINDINGS WITH NO THREAD` block follows naming the shortfall, except
           where a count-first open section supplied the total, as stated below, or where open
           entries standing for an earlier round's threads account for it, as stated below. Two
           things overstate that shortfall rather than hiding it, so it is confirmed against the
           body rather than acted on from the number. One is a thread past the hundred
           `reviewThreads` reads, which would have counted in `M`, and `threads=` carries the
           trailing `+` saying the page was cut. The other is a finding an earlier round raised
           that the total still counts and no open entry links, its thread being none this round
           opened. An open entry whose title links an earlier round's thread, and which links
           none this round opened, is counted beside `M` before the two are compared, since the
           total counts what is still open rather than what this round raised. So `T` above `M`
           can print no block, and where a block prints it names how many it counted that way.
           A withheld finding is owed the triage a suppressed one is. No exit code rides on it,
           the same as `suppressed=` and `cr_outside_diff=`, since what an unread finding says is
           the reader's to judge.
           Where the preamble states no total, the count opening the open-findings section is
           `T`, since that format's later revision writes the bare zero line only for a zero.
           Each entry an open section lists links its thread, and an entry whose title links no
           thread on the pull request is counted in that same block whatever the totals say. So
           is each entry an `Open (N)` section counts beyond its linked entries, the bullets at its
           margin carrying a link this script reads, outside any block collapsed within it.
           Where the open-findings section supplied `T`, the block is measured against the
           entries it lists with a link rather than against `M`, so an entry linking an earlier
           round's thread is no finding without one, while a finding that section's count names
           and no entry links is still counted. Entries are counted rather than links, so a
           back-reference beside an entry's own link covers no other entry.
           A round can also list its findings in prose under a `Review findings:` label, each
           entry naming the files it concerns in code spans. An entry naming no file a thread of
           that round is on is a finding with no thread, so where those entries outnumber the
           shortfall above, the block counts them instead.
           Ten bodies in that format have been read here, so what follows describes those rather
           than the format in general. None carried a `Suppressed comments` heading, which is
           why this field rather than `suppressed=` is what finds a withheld finding in one.
           None stated its file coverage the first format's way, in either the bulleted spelling
           or the bare one, so the `fleet-review` marker the fleet's own carried `fleet-code-review`
           instructions ask for is the only coverage statement this format has been seen to make.
           Five of the ten carry it, on heads carrying those same instructions, so its absence is
           intermittent rather than structural. Both absences measured on one drive were
           re-reviews of a delta of one line and of four files, which is why a round stating none
           carries an earlier round's statement forward rather than blocking. A re-request is not
           the remedy the block named either: four were measured and one produced the marker, so
           it clears by chance rather than by asking.
           A later body in that format that
           does collapse a suppressed section is read by `suppressed=`, that heading being vetted
           and read wherever it sits. Whether the shortfall counts those findings a second time
           turns on whether the stated total includes them, which no body read here says, so treat
           `suppressed=2` beside a shortfall of 2 as either two findings or four until the body
           settles it. A body stating coverage as a bare `**Files reviewed:**` line would still go
           unread, the reader requiring a bullet where this format states its metadata without
           one, which is why `effort=` had to drop that requirement. Ten rounds have stated it
           that way none of the time, so this is a gap rather than the reason a round reads as
           unstated, and `Files reviewed` sitting in the vetted labels does nothing about it
           since the line never reaches that comparison.
           Nothing reports either contradiction, the vetted lists reporting an unvetted marker
           rather than a marker turning up somewhere new.
           Two quotation shapes carry a column-0 `<details>` past both masks, a multi-line HTML
           comment and a raw `<pre>` block, and either ends the preamble and prints `?`. The
           indented code block is a third, answered by the pattern's own bound rather than a mask.
  reply    Answer one thread selected by its text, and resolve it on request. Exists
           because the hand-run form keeps failing the same way: a node id typed into a
           mutation, which resolves globally and so writes to a real thread somewhere
           rather than failing. This takes a pull request number and words from the
           finding, queries the id itself, and offers no argument an id fits in. A finding
           posted twice on one file, threads whose opening comments read the same, is
           answered as one under --all-identical. Exit 0 = done, 60 = no thread matched,
           61 = more than one did and the flag did not apply, 62 = the reply returned
           no comment url so nothing was resolved, 63 = the resolve did not report the
           thread resolved, 64 = the write scope could not be established or excludes the target.
  wait     Request a review where none is outstanding, then poll until Copilot's review lands on the
           current head, then print the digest. The auto-request is skipped once a review already
           covers the head, once Copilot has already answered outside a formal review, or once one
           is already in the pending request set, so calling `wait` again on the same PR never
           double-requests. It is also skipped under 46's and 47's quota readings below, since a
           request into a reached limit spends quota and returns the same refusal. It is skipped too
           on a pull request into a branch other than the default once Copilot has reviewed it at
           all, since a fix push there is covered by an attested local pass: an attested head is
           covered at once, so the wait polls its checks instead only while they can still move
           its merge, as 30 below lists, and one with no attestation exits 49 naming the `attest`
           step.
           --request asks for a round anyway. A pull request into the default branch, a promotion
           among them, a pull request Copilot has not reviewed yet, and one with a partial on record
           or a review history past the window are requested as before. The comment also carries the
           findings count the pass recorded, shown and not gated. It reads the Copilot reviewer's
           bot id from the repository's own most recently updated PRs rather than a fixed id: the
           last HISTORY_PRS, widened once to HISTORY_PRS_WIDE where that narrow window carries no
           Copilot activity at all, since an outage that outlasts HISTORY_PRS PRs would otherwise
           empty it on every call for as long as the outage runs. Requests nothing (falling back to
           polling only) where both windows come up empty, since a repository with no Copilot review
           in either has nothing to read the id from and a fabricated one is never an option. The
           loop runs in-process, so a 45-minute wait costs one agent turn, not 90.
           Exit 0 = review present, or on a held head an attested local pass with its required
           checks settled as far as the rollup window reads them, with the merge reading CLEAN,
           UNSTABLE, or HAS_HOOKS, or with the pull request closed, 30 = still pending at timeout
           (pending is not failure), on a held head a required check still settling, any check while
           no required one has posted, none posted on a BLOCKED, BEHIND, or DRAFT merge, or the
           merge still UNKNOWN, printed as `status=CHECKS_PENDING`, or a round requested while the
           held poll ran, printed as `status=PENDING` unless 40's outside answer or 46's or 47's
           quota reading outranks it,
           40 = Copilot answered outside a formal review, so read the printed body.
           40 reports the shape of that answer and reads nothing of its cause: an answer
           carrying no commit covers no head, so the wait ends and the reader decides.
           41 = the review carrying the head says it did not review, so it covers nothing.
           42, 43, 45, and 50 = the review landed and `status`'s readings apply to it.
           A wait ending on a round that covered half the diff, or on output nothing here can
           read, has ended on something other than a review of this pull request, and one ending
           on a 50 has ended on a review whose finding no count carries. 50 outranks 44, so on a
           head carrying a 50 a stuck check shows only in the digest. A 42 can
           be decided by an earlier round whose partial coverage carries to a head stating none,
           in which case the round the wait ended on is not the round the code came from.
           44 = the review loop closed, the merge reads BLOCKED, and a required check is in a shape
           no wait clears: queued with nothing acting on it, expected and never posted, running far
           past what the job costs, or failed, though on a held head a required check running long
           is still polled, as 30 states. A check merely still running normally is not this and
           exits 0, or on a held head is polled while it can still move the merge. Neither is a
           stuck check its rollup node reads as not required, since the rollup carries checks no
           ruleset requires, nor any stuck check on a merge that is not BLOCKED. The digest reports
           the check in every case, so a shape outside 44 is still named rather than lost.
           46 = the newest Copilot review on the pull request, on this head or an earlier one, is a
           refusal naming the account quota, or one saying only that it encountered an error,
           printed above under COPILOT REFUSED THIS ROUND. The weekly rate limit posts that error
           body and writes its cause to the reviewer's own Actions run, so that run's job log is
           read: a logged rate limit reports as the quota with its reset time, and anything else, an
           unreadable log included, as a possible quota hit. Either way no request is sent, and a
           request already pending after a refusal on an earlier head is still polled for. That is
           an account-level state a re-request or a further wait does not clear, unlike 41's other
           causes (a file count over the limit, cleared by splitting the pull request), so it is its
           own code rather than folded into 41: proceed on the other reviewers' coverage instead of
           retrying.
           47 = this pull request's current head carries no Copilot activity of its own, and the
           reviewer's own most recent activity found in the repository, a review or comment on any
           other pull request, is that same account-quota refusal with nothing having answered it
           since. A refusal on this pull request itself is 46 instead. The poll that would otherwise
           have run is skipped for this reason, printed as a `note:` line before the digest, rather
           than spent finding the same account state out a call late, and no request is sent, while
           a request already pending is still polled for. Pass --ignore-quota-signal to request and
           poll --timeout anyway once the quota is believed to have reset. 46 is read from this pull
           request's own reviews and always takes priority over 47, so a genuine 0/40/41/42/43/45 on
           this pull request outranks 47 whenever both would otherwise apply, and so does a 50, and
           on a held head covered by an attested local pass a 30 or a 44.
           A pending request remains pending until a review, an answer, or the timeout. GitHub's
           effort-labeled review lifecycle does not always emit `copilot_work_started`, so that
           event is not evidence that distinguishes queued work from abandoned work.
           48 = the auto-request returned success and recorded nothing on the pull request,
           neither a pending reviewer nor a review-request event. Observed once, while the
           requesting account's Copilot allowance was exhausted, where no refusal was posted to
           read and clearing the set and requesting again changed nothing. It is decided one
           poll interval after the request, and the rest of the poll is skipped, since no
           request exists to answer. It cannot meet 46 or 47, since neither sends a request,
           and ranks under 0/40/41/42/43/44/45/50. `status` cannot report it, since a request that
           recorded nothing leaves nothing for a later read to find.
           49 = no Copilot round covers this head and none was requested, the pull request merging
           into a branch other than the default after Copilot's first round, and no attestation
           vouches for the head. Run the local strict review, record it, push, and run `attest`, or
           pass --request.
           64 = the write scope could not be established or excludes the target, checked before
           the auto-request or any poll, so a cross-owner target reads and writes nothing here.

Reading is the bulk of this and the writing commands are a trade rather than a free win. A
mutation spelled as a `gh` command in a shell is read by the gh-write-guard PreToolUse hook, and
one this script performs is not. The script removes the guarded failure at its source: every node
id comes from a live query in the same run, and the owner boundary is enforced in-process. See
.github/copilot-instructions.md for the runbook, and GOVERNANCE.md "Repository Boundaries and
Write Safety" for the rules these commands enforce.
"""

from __future__ import annotations

import argparse
import fnmatch
import functools
import html
import io
import itertools
import json
import os
import re
import subprocess
import sys
import tarfile
import time
import unicodedata
from collections.abc import Callable, Iterable
from datetime import UTC, datetime
from pathlib import Path

# The directory this script sits in, and the anchor every scope answer is read from.
# One value serves the owner probe and the refusal naming it, so the two cannot disagree.
HERE = Path(__file__).resolve().parent

REVIEWER = "copilot-pull-request-reviewer"
# Other review bots this repository has trialed alongside Copilot.
# Tracked at the identity level only, login and commit oid, never body prose, except where a reader below names one explicitly.
# Each format read here is its own reader, and doing that well is a separate task per bot.
CODERABBIT_LOGIN = "coderabbitai"
QODO_LOGIN = "qodo-free-for-open-source-projects"
# Named rather than inlined at each of their own readers below, so a login rename updates one spelling instead of silently leaving a hardcoded copy matching nothing.
OTHER_REVIEWERS = (CODERABBIT_LOGIN, QODO_LOGIN)
KNOWN_REVIEWERS = (REVIEWER, *OTHER_REVIEWERS)

# A check dispatched to a runner and not begun, in the spellings a CheckRun carries.
# PENDING here means dispatched and not begun, which a StatusContext's means the opposite of.
# That one is translated away in check_nodes, where the node shape is still known.
# Reading QUEUED alone scores the rest of these as running.
NOT_STARTED = frozenset({"QUEUED", "WAITING", "PENDING", "REQUESTED"})
# A required status nothing has posted, which is a StatusContext's state and only ever that.
# It is not a starved job and must not borrow that remedy, since no runner is owed it at all.
# Left out of both sets it reaches the conclusion branch, where it reports as a red check.
NOT_POSTED = frozenset({"EXPECTED"})
# What counts as a check having passed, rather than as one still deciding or failed.
# SKIPPED and NEUTRAL are passes, since the fleet aggregator pattern skips the conditional jobs.
# Treating a skip as unfinished reports four blockers on every green pull request here.
CHECK_OK = frozenset({"SUCCESS", "SKIPPED", "NEUTRAL"})
# How long a queued check is simply a check starting.
# Observed pickup on a healthy run here is one to two minutes.
# This is the pickup grace's five minutes for the same reason that one is.
CHECK_GRACE = 300
# How long a running check runs before its duration is worth reporting.
# Generous on purpose, since this repository's lint job legitimately takes eleven minutes.
# A fleet repository building and testing .NET takes longer still.
# A threshold that flags those trains the reader to skim the genuinely hung job too.
CHECK_STALL = 1800

# A review body can carry a collapsed block of findings withheld from the inline threads.
# Those appear nowhere in `reviewThreads`, so polling threads alone reports a clean pass.
# The alternation is the runbook's, since the heading wording has changed once already.
# Matching one phrasing alone reports zero on a review that has them.
SUPPRESSED = re.compile(r"Suppressed comments|low confidence", re.IGNORECASE)
# The second overview format's own section for a finding against code the pull request did not change.
# Anchored to the opening of its line, past whatever markup a heading wears, rather than matched anywhere in it.
# The wording is ordinary English, so a body-wide match reads a sentence such as `restores a guard previously missed (2) rounds ago` as this section.
# `marker_blocks` accepts a line on a count alone, which is what lets that sentence through, so the anchor is what separates a heading from prose naming one.
# The opener admits heading markup and anything outside ASCII, and nothing else.
# An emoji is what this format already puts on its verdict headings, and a spelling that admitted no such character read a heading wearing one as no section at all.
# That failure is silent rather than loud, the heading itself being vetted and `normal` dropping the emoji before the comparison, so the digest reports every shape read while the findings reach nobody.
# Widened to any non-word character instead, ordinary punctuation opened the run and a sentence such as `- Previously missed (2) findings were re-raised` became the section, printing the prose around it as the finding.
# Punctuation is what prose opens with and markup is what a heading opens with, so the class is the markup rather than the complement of a word.
# The whitespace is spelled out rather than taken from `\s`, which is Unicode-aware and so overlaps the non-ASCII alternative on every non-ASCII space.
# Two alternatives able to consume the same character make the run ambiguous, and the parse then doubles per character of it on a line that does not go on to say this, which is a hang rather than a slow read.
# A body carrying 30 non-breaking spaces in a line took minutes, and every alternative here now opens on a character the others cannot.
# An emphasis marker is admitted only where no space follows it, which is what tells `**Previously missed (1)**` from the list item `* Previously missed (2) findings`.
# A bullet is the one prose opener that is also heading markup, and `marker_blocks` accepts a line on a count alone, so the list item read as the section and printed the prose around it as findings.
# One character per iteration rather than a run of them, since a repeated group that can split a run of stars more than one way is ambiguous and parses exponentially in it.
# What this still admits is non-ASCII punctuation, an em dash or a typographic quote opening a sentence, which reads that sentence as the section.
# It is narrower than the spelling before it, which admitted ASCII punctuation as well, and the remedy for the rest is a class of non-ASCII symbols rather than the whole of non-ASCII.
PREVIOUSLY_MISSED = re.compile(
    r"^(?:</?[A-Za-z][^>]*>|[ \t>#]|[*_](?![ \t])|[^\x00-\x7f])*Previously missed\b",
    re.IGNORECASE,
)
# CodeRabbit's own equivalent, collapsed into the review body like `SUPPRESSED` rather than raised as an inline comment.
CR_OUTSIDE_DIFF = re.compile(r"Outside diff range comments", re.IGNORECASE)
# A refusal declines the round as a formal review carrying the head and no threads.
# That is the clean pass byte for byte, so every coverage check passes over a round that never ran.
# The alternation is the runbook's for the same reason the one above is.
# One phrasing is one rewording away from reading a refusal as a review.
# The dot covers the apostrophe in the typographic spelling and the ASCII one alike.
# It also keeps the published filter usable inside single quotes, which neither survives.
REFUSAL = re.compile(
    r"wasn.t able to review|was not able to review|unable to review", re.IGNORECASE
)
# The account-level cause among the refusal wordings, as opposed to a per-pull-request one like the file count.
# Read against `refusal_of`'s own return rather than the raw body.
# That way the exemptions built for that anchor, the opening line and the fenced quotation, protect this reading too instead of needing their own.
# This script's own corpus and this file both quote the sentence below its overview, same as the refusal wording itself does.
# Observed once here: "Copilot was unable to review this pull request because the user who requested the review has reached their quota limit."
QUOTA = re.compile(r"reached (?:their|its|his|her|your|my) quota limit", re.IGNORECASE)
ERROR_REFUSAL = re.compile(r"encountered an error", re.IGNORECASE)
COPILOT_RUN_PATH = "dynamic/agents/copilot-pull-request-reviewer"
RUN_ERROR_TYPE = re.compile(r"errorType: '([a-z_]+)'")
RUN_RATE_LIMIT = re.compile(
    r"(You.ve reached your [^\n]*?rate limit\.[^\n]*?)(?= or switch| Learn More|$)", re.MULTILINE
)
ANSI = re.compile(r"\x1b\[[0-9;]*m")
ISO_STAMP = re.compile(r"\d{4}-\d\d-\d\dT\d\d:\d\d:\d\dZ")
# A structural marker rather than prose, so reading it needs no per-bot wording model the way `QUOTA` above needs one for Copilot's free-text refusal.
# Observed on CodeRabbit, a plain PR comment rather than a formal review: "<!-- This is an auto-generated comment: rate limited by coderabbit.ai -->".
# The service name is captured rather than assumed.
# A second bot using the same auto-generated-comment convention is read without a new pattern, and one that does not use it stays unread rather than guessed at.
RATE_LIMITED = re.compile(r"auto-generated comment:\s*rate limited by\s*(\S+?)\s*-->")
# Anchored to the `<h3>` tag rather than a bare substring, since prose elsewhere can mention the phrase without being the comment it names.
QODO_REVIEW_HEADING = re.compile(r"<h3>\s*Code Review by Qodo\s*</h3>", re.IGNORECASE)
# Only the numbered heading counts as a finding, told apart from Qodo's own nested sub-summaries by starting with a number and a period (`1.`, `2.`, ...).
QODO_FINDING = re.compile(r"\s*\d+\.\s")
# A finding's own title can quote `Resolved`/`Dismissed` without carrying the badge, so the glyph is required rather than just the word.
# Escaped (U+2713, U+2717) rather than typed literally, per the repository's ASCII charset rule.
QODO_BADGE = re.compile(r"<code>[^<]*(?:\u2713 Resolved|\u2717 Dismissed)[^<]*</code>")
# A round states how much of the diff it read on a line of its own.
# A round that read part of it is the clean pass elsewhere, same commit and threads and digest.
# Five such rounds landed across three merged pull requests here.
# One of them read 2 of 3 changed files across both its rounds and merged.
# This is the third instance of the shape the two patterns above answer.
# It is also the only one nothing was reading.
# The line is anchored at its start rather than matched body-wide, both spellings being structural.
# Over 333 review bodies every coverage statement opens its line and not one sits mid-sentence.
# 272 of them open with the reviewer's own name and 32 are the `Review details` bullet.
# A body-wide match reports the pull request adding this check as a partial round.
# That is the false positive the suppressed matcher and the refusal matcher have each had once.
# The cost of the anchor is named rather than hidden.
# A wording that moves the statement off the line start reads as no statement at all.
# Two openers rather than one alternation, because the text each needs beside it differs.
# The bullet's own label is the marker, so it is a coverage line whatever follows the counts.
# Requiring the trailing words there made a bullet that drops them read as no statement at all.
# The sentence opener is the reviewer's name, which prose also opens a line with.
# That one keeps the text requirement, since the name alone does not say the line states coverage.
COVERAGE_BULLET = re.compile(r"\s*[-*]\s*\*\*Files reviewed:", re.IGNORECASE)
COVERAGE_SENTENCE = re.compile(r"\s*Copilot\b", re.IGNORECASE)
# A blockquote marker (`>` per line, Markdown's own quoting convention) sits in front of every line CodeRabbit wraps its outside-diff section in.
# `HEADING`'s `\s*` does not match `>`, so a heading under one reads as prose without this stripped first.
# Read by two callers for opposite purposes, which is why it is defined here rather than beside either.
# `marker_blocks` strips it to find a heading underneath, and `coverage_statements` refuses the line outright.
# A quoted marker is a quotation whichever convention quotes it, and the coverage reader has no use for what a blockquote carries.
BLOCKQUOTE = re.compile(r"^>+\s?")
# The severity the second overview format now renders as a theme-switched `picture` element.
# Its only textual carrier is the `img` `alt`, every other part being a URL that changes with the icon set.
# The set is versioned in those URLs and has already been bumped once, so matching them would pin this reader to one revision.
# Reduced to the `alt` text rather than dropped, since the severity is the half of a finding entry worth reading.
# Either quoting is read, both being valid HTML and the attribute's own spelling being exactly the kind of presentation this reader exists not to depend on.
# Read in one only, a change of quoting leaves the badge unreduced, which prints two icon URLs per finding and can report the entry as a shape nothing here knows.
# Neither gap may cross a `picture` tag, or one match spans from the first element to a later element's `img`.
# An entry whose own `img` carries no `alt`, which the icon set drifting is exactly how to get, then pairs with the next entry's.
# Everything between, that entry's title and its whole finding text, is replaced by the later severity and lost from the only place these findings appear.
# Bounded, such an entry keeps its markup and reads as an unrecognized shape, which is the failure that reports itself.
SEVERITY_BADGE = re.compile(
    r"<picture\b[^>]*>(?:(?!</?picture\b).)*?"
    r"<img\b[^>]*\salt=(?:\"([^\"]*)\"|'([^']*)')[^>]*>"
    r"(?:(?!</?picture\b).)*?</picture>",
    re.DOTALL | re.IGNORECASE,
)
# What is left of a finding entry's own heading once the badge above is reduced to its text.
# A finding entry cannot be vetted the way a section heading is, its title being different on every finding.
# The badge is therefore what identifies the shape, and the title after it is free text by design.
# The severity alone does not separate the two, a section being free to name one in its own heading.
# Every vetted section heading here states its count and a finding's title states none, so the absence of one is the other half of the test.
# Without it a section this has no reader for is vetted on its first word, which turns a shape that reports itself into one that is silently passed over.
SEVERITY_TEXT = re.compile(r"(?:critical|high|medium|low)\s+severity\b", re.IGNORECASE)
# The count as `normal` leaves it rather than as the heading wrote it.
# `normal` rewrites every `(\d+)` to `(N)` before a marker is ever compared, so a reader keyed on the digits finds none and reads every section heading as carrying no count.
NORMALIZED_COUNT = re.compile(r"\((?:\d+|N)\)")
# Ask the reviewer for a stable coverage shape instead of adapting only to changing prose.
# Keep the prose readers for reviews made before the marker shipped.
# Searched within its line rather than matched as the whole of one, because the reviewer writes it both ways.
# It carries no whitespace of its own at either end, which a search has no use for and pays for: an unanchored run of `\s*` is quadratic in the whitespace run it sits in.
# Left in place from the spelling that matched a whole line, a line of 16000 spaces took half a second and 65000 took nine, where the marker itself is what a search is looking for.
# The instruction it is emitted from asks only that the body end with it, and one of the two overview formats ends a prose sentence with it instead.
# Anchored to the whole line, that format read as a round stating no coverage at all, over a body carrying the counts in plain sight.
# The anchor was never what made the marker trustworthy: an HTML comment renders invisibly, so a body that wants one seen has to quote it, and a quotation is refused before this is ever consulted.
# The digit runs are matched unbounded and judged in `read_coverage` rather than bounded here, which is the difference between a line this cannot believe and a line that is not a statement at all.
# Bounded here, a marker stating a count longer than any review states stopped matching, so the line left the gate silently and a round carrying one read as full on whatever else it said.
# `int` also raises on a run past 4300 digits, so the judgment has to happen somewhere, and the place that can answer `unvetted` is the reader rather than the pattern.
FLEET_REVIEW = re.compile(
    r"<!--\s*fleet-review:\s*reviewed=(\d+)\s+changed=(\d+)\s+findings=(\d+)\s*-->",
    re.IGNORECASE,
)
# The count pair itself, in the two spellings the corpus carries.
# The comment tail one of them ends on is deliberately not part of the unit.
# It says how many comments the round raised, which is not coverage.
# A fifth wording of it would fail every merge over a sentence ending read correctly.
# The plural is optional, since a one-file round reading `1 changed file` means what it says.
# Blocking on that is the cry-wolf case, a fleet-wide stop over a grammatical agreement.
# The bullet's trailing words are optional for the reason its label alone identifies the line.
# Detection and parsing disagreeing there turned a readable `4/4` into a block on a readable line.
# What is left blocking is a bullet carrying no counts, which genuinely states no coverage.
COVERAGE_COUNTS = re.compile(
    r"reviewed\s+(\d+)\s+out of\s+(\d+)\s+changed files?"
    r"|\*\*Files reviewed:\*\*\s*(\d+)\s*/\s*(\d+)(?:\s+changed files?)?",
    re.IGNORECASE,
)
LOOSE_COUNTS = re.compile(r"reviewed\D{0,40}?(\d+)\s*(?:out of|of|/)\s*(\d+)", re.IGNORECASE)
EMPHASIS_MARK = re.compile(r"[*_`]")
HTML_TAG = re.compile(
    r"</?[A-Za-z][\w-]*(?:[ \t]+[\w:-]+(?:[ \t]*=[ \t]*(?:\"[^\"\n]*\"|'[^'\n]*'|[^\s\"'<>]+))?)*[ \t]*/?>"
)
MARKER_NAME = re.compile(r"fleet[\W_]{0,3}review", re.IGNORECASE)
MARKER_NAME_TIGHT = re.compile(r"fleet\W?review", re.IGNORECASE)
LINE_TAG = re.compile(r"</?[A-Za-z][A-Za-z0-9-]*(?:\s[^<>\n]*)?/?>")
BARE_TAG = re.compile(r"<(?!!--)[^<>]*>")
MARKER_REACH = 300
# A fenced block is a quotation rather than a statement, and 131 of those bodies carry one.
# This change puts both spellings into the source and the runbook, so a review of it quotes them.
# A quoted count read as this round's own is a coverage figure nobody stated.
FENCE = re.compile(
    r"^ {0,3}(?:(`{3,})[^`\n]*|(~{3,})[^\n]*)(?:\n.*?)??"
    r"(?:(?P<close>\n {0,3}(?:\1`*|\2~*)[ \t]*\r?(?=\n|\Z))|\Z)",
    re.DOTALL | re.MULTILINE,
)


def strip_fences(
    body: str, repl: Callable[[re.Match[str]], str] = lambda m: "", to_end: bool = False
) -> str:
    """The body with each fenced block replaced by `repl`, the unclosed one kept unless `to_end`.

    Markdown runs an unclosed fence to the end of the body, so for the coverage reader everything
    after one is a quotation, and masking it there reads as no coverage, which blocks. Every other
    reader looks for something that blocks, such as a suppressed finding or an unrecognized shape,
    and masking the rest of the body there would hide it, so those keep reading it.
    """
    return FENCE.sub(
        lambda m: repl(m) if to_end or m.group("close") is not None else m.group(), body
    )


# An inline code span is a quotation for the same reason a fenced block is.
# A reviewer naming `<summary>` in prose was read as opening one.
# Bounded to a paragraph, and an escaped tick opens nothing, since either masks a section.
# Only the opener takes that guard, a backslash being literal inside a span rather than an escape.
# Replaced by a space, since deleting joins the sides into a tag the body never carried.
CODE_SPAN = re.compile(r"(?<![\\`])(`+)(?!`)(?:[^\n]|\n(?!\s*\n))*?(?<!`)\1(?!`)")


# The round's own file summary table, whose first column is a path and whose second is prose.
# The header row is the marker rather than the `<details>` around it.
# One measured round carries the table with no `<summary>`, and two summary spellings wrap it.
# Over the same 348 bodies, 91 carry a table and every table row in the corpus belongs to one.
# A row outside this shape is a shape nothing has seen rather than a table read wrongly.
TABLE_HEADER = re.compile(r"\s*\|\s*File\s*\|", re.IGNORECASE)
# The alignment row under the header, which is punctuation rather than a file.
TABLE_RULE = re.compile(r"\s*\|[\s:|-]+\|\s*$")
TABLE_ROW = re.compile(r"\s*\|([^|]*)\|")
TABLE_GLOB = frozenset("*?")
TABLE_GAP = " ... "
TABLE_SPAN = re.compile(r"`([^`]+)`")
TABLE_SEPARATORS = re.compile(r"[\s,]*")
# The readings a round's coverage carries, worst first.
# A head carries more than one round only through a re-request.
# Where two disagree, the one naming files it did not read is the one to answer.
# `UNSTATED` sits last rather than beside the failures, being the absence of a statement.
# A round that did state full coverage settles the question over one that stated nothing.
UNVETTED, PARTIAL, FULL, UNSTATED = "unvetted", "partial", "full", "unstated"
# Not a state a round reports, but the reading that carries an earlier round's forward.
CARRIED = "carried"
TABLE = "table"
LOCAL = "local"
NO_HEAD_TABLE = "no round covering the head carries a file table of its own"
SEVERITY = (UNVETTED, PARTIAL, FULL, UNSTATED)
# Upper-case for the two that block a merge, for the reason `review_on_head=NO` is upper-case.
# `unstated` rather than `unknown`, since a body carrying no count is a shape this knows.
# What this script does not know is the separate `shapes` field, and one word for both hides it.
# The constant carries that name too, so no reader has to map a name here onto another word.
COVERAGE_FIELD = {
    UNVETTED: "UNVETTED",
    PARTIAL: "PARTIAL",
    FULL: "full",
    UNSTATED: "unstated",
    TABLE: "table",
    LOCAL: "local",
}

# Every structural marker the reviewer's own bodies carry, measured over the same 333.
# A body is read for these rather than trusted, because every reader below keys on one of them.
# A heading this script has no spelling for is a section it will not find, reported as absent.
# That is the shape of all three failures already on record here, each caught after it landed.
# The lists are small because the output is regular: 10 headings, 13 summaries and 4 labels.
# Two overview formats are carried rather than one, the second arriving after that measurement.
# Its own markers are listed beside the first's, since a round in either format can land next.
# Counts are normalized to `(N)` and non-ASCII is dropped before comparing, and `unvetted` folds letter case at the comparison itself.
# The verdict headings carry a colored circle, so the emoji is what would drift most cheaply.
# Dropping it also keeps this file inside the charset rule that governs the repository.
CLEAN_VERDICTS = {"### Approval recommended", "### Ready to approve"}
FLAGGING_VERDICTS = {
    "### Changes recommended",
    "### Needs a closer look",
    "### Not ready to approve",
    "### Human review recommended",
}
VETTED_HEADINGS = (
    {
        "## Pull request overview",
        "## Copilot review overview",
        "### Reviewed changes",
        "### Suppressed comments (N)",
    }
    | CLEAN_VERDICTS
    | FLAGGING_VERDICTS
)
NARRATIVE_SUMMARY = "What changed in this PR"
TOTAL, OPEN, RESOLVED = "total", "open", "resolved"
OVERVIEW_SECTIONS = {
    "Open (N)": OPEN,
    "(N) open findings": TOTAL,
    "(N) open finding": TOTAL,
    "Resolved since last review (N)": RESOLVED,
    "(N) resolved since last review": RESOLVED,
}
VETTED_SUMMARIES = set(OVERVIEW_SECTIONS) | {
    "Pull request overview",
    NARRATIVE_SUMMARY,
    "Show a summary per file",
    "File summaries",
    "Review details",
    "Suppressed comments (N)",
    "Comments suppressed due to low confidence (N)",
    "Previously missed (N)",
}
VETTED_LABELS = {"Files reviewed", "Comments generated", "Review effort level", "Review effort"}
# Inline emphasis around a marker is presentation rather than identity.
# The second overview format wraps every `<summary>` in `<strong>` where the first wrapped none.
# Vetting the tags along with the text blocks on markup changing under a section that has not.
# That is the same cheap drift the verdict headings already drop their colored circle for.
EMPHASIS = re.compile(r"</?(?:strong|b|em|i)>", re.IGNORECASE)
MARKDOWN_HEADING = re.compile(r"\s*#{1,6}\s")
# The `Review details` metadata bullets, of which the coverage line is one.
# The bullet stays required, so the second format's bare bold metadata lines are not read here.
# Dropping it reads that format's own `**Changes:**` prose label as metadata and blocks on it.
# Each bare line that carries a reading has its own pattern below instead.
LABEL_LINE = re.compile(r"\s*[-*]\s+\*\*([^*]+):\*\*")
# Two spellings, one per overview format.
# The first states it as a `Review details` bullet.
# The second has no `Review details` section and states it as a bare bold line.
# Reading the first alone printed `effort=unknown` over a body stating the level outright.
# A later revision opens the second on an emoji, a non-ASCII run excluding whitespace so `\s*` cannot overlap it.
EFFORT_LINE = re.compile(
    r"\s*(?:[-*]\s+)?(?:[^\x00-\x7f\s]+\s*)?\*\*Review effort(?: level)?:\*\*\s*"
    r"(?:(Default)\s*\(\s*(Lite|Balanced|Max)\s*\)|(Lite|Balanced|Max))\s*$",
    re.IGNORECASE,
)
# The second overview format states its own version in the marker its body opens with.
# Keyed on that marker rather than on the heading wording, a version being what it states.
# Anchored to a line of its own, since the marker is an HTML comment and renders invisibly.
# A body naming it mid-sentence is a body about the format rather than one written in it.
# A writer who wants it seen at all has to quote it, which is why a code span is masked first.
# Bounded to three spaces of indent, which is the rule `FENCE` reads a fence by.
# A fourth makes the line a code block, and so a quotation the fence guard never sees.
CCR_OVERVIEW = re.compile(r"^ {0,3}<!--\s*ccr-overview-v2\s*-->\s*$", re.MULTILINE)
# That format's own finding total, which the first format states nowhere.
# Line-anchored for the reason a coverage line is, prose being able to carry the words mid-line.
# Bounded to three spaces for the reason the marker above is, a fourth making it a code block.
# The largest total wins, so a quoted number otherwise beat the round's own.
# `None` is a total of zero spelled in words, which a round raising nothing states.
# Read as no total at all it printed `?`, and this file's own rule is that the two differ.
# The rest of the line is taken whole rather than its first number.
# The format states a count per severity, so a round stating `2 <medium> . 3 <low>` raised five.
CCR_FINDINGS = re.compile(r"^ {0,3}\*\*Findings:\*\*(.*)$", re.MULTILINE)
CCR_OPEN_FINDINGS = re.compile(
    r"^ {0,3}\*\*(\d+) open findings?\*\*[ \t]*\r?$", re.MULTILINE | re.IGNORECASE
)
# The markup the format writes a count beside, replaced by one sentinel before a count is read.
# Masked rather than matched around, because a badge carries digits of its own in its attributes.
# `2 <img src="b.svg" width="62" height="18">` read `18` as a severity count on that reasoning.
# Dropping the markup outright would join the digits either side of it into one number instead.
# The tag alternative is bounded by `[^<>]`, so it cannot span from one tag into another.
# An unmatched `<` earlier in the line otherwise paired with a later tag's closing `>`.
# Every count between the two was swallowed with it.
# It opens on a name character for the same reason, a `<` followed by a space being prose.
# The link alternative tolerates one level of nested parentheses and brackets, a URL carrying a
# `)` of its own having otherwise ended the match early and left its anchor id on the line.
# A reference-style link is its own alternative, its label being a number as often as a word.
COUNT_MARKUP = re.compile(
    r"</?[A-Za-z][^<>]*>"
    r"|!?\[[^\[\]]*(?:\[[^\[\]]*\][^\[\]]*)*\]\([^()]*(?:\([^()]*\)[^()]*)*\)"
    r"|!?\[[^\[\]]*\]\[[^\[\]]*\]"
)
# What opens a block of its own, and so ends a blockquote rather than continuing it.
# Markdown's lazy continuation carries paragraph text only, which of the three coverage spellings is the reviewer's own sentence alone.
# The bullet spelling and the marker each open a block, so each is the round's own line again however the lines above it are quoted.
# Read as continuations they reported `unstated` over a round stating full coverage, which blocks a merge on a satisfied item.
# A bare `<` is not one of these, an inline tag opening a line of prose leaving that line paragraph text.
# A tag standing alone is not one either, though it looks like it should be: only some tag names open a block that can interrupt a paragraph, and a lone `</span>` is not among them, so accepting any of them read a quoted line as this round's own.
# No coverage spelling opens with a tag in any case, the bullet opening on its marker and the sentence on the reviewer's name, so the alternative bought nothing and is dropped rather than narrowed to a list of tag names.
# The caller also holds the four-space bound, an indented line being a code block, which cannot interrupt a paragraph and so cannot end a quotation either.
STARTS_BLOCK = re.compile(r"[-*+]\s|#{1,6}\s|<!--")
# A control character, so no body can carry one of its own and be read as having a count here.
MARKUP_MASK = "\x00"
# Every integer the markup did not swallow, which is how this format states each severity.
# Bounded to four digits, because a review states tens of findings and never thousands.
# A longer run is an identifier, so the mask is not the only thing saying an anchor id is not one.
# That bound is also what keeps `int` from being handed a digit run long enough to raise.
# A digit touching a `.` is part of a decimal, so `confidence < 0.5` contributes no count.
COUNT_LOOSE = re.compile(r"(?<![\d.])\d{1,4}(?![\d.])")
# The digits a count may carry before it is a run to refuse rather than a number to read.
# `int` raises above 4300 of them, so every conversion of a matched run goes through `counted` rather than straight to `int`.
# Bounding the pattern instead would leave the line unmatched, which reads as no statement at all rather than as one this cannot believe.
# Four is what `COUNT_LOOSE` above already holds its own runs to, a review stating tens of files and never thousands.
COUNT_DIGITS = 4
# A total of zero spelled in words, anchored to where the line states its total.
# Read as a bare substring instead, `see nonetheless the summary below` stated a zero.
# Emphasis markers are skipped, the format writing the word italicized as often as bare.
# Bounded by a lookahead rather than by `\b`, since `_` is a word character.
# `_None_` then ended on no boundary at all, reading an italicized stated zero as none stated.
STATED_NONE = re.compile(r"^[\s_*]*none(?![a-z0-9])", re.IGNORECASE)
# The first section opener on a line of its own, which is where the overview preamble ends.
# Read case-insensitively as every other tag reader in this file is, since a body spelling it
# `<DETAILS>` otherwise never ends the preamble and a section's own total becomes the round's.
# Anchored to a line for the reason the marker and the total above are.
# Bounded as `DETAILS_TAG`'s opener is, so `<details/>` ends the preamble and `<detailsfoo>` does not.
# Matched anywhere instead, a round naming `<details>` in its overview prose ended the preamble.
# That threw its stated total away, printing `?` and no shortfall over a round withholding findings.
DETAILS_OPEN = re.compile(r"^ {0,3}<details(?=[ \t\n\f\r]|/?>)", re.IGNORECASE | re.MULTILINE)
# A login that reads as this reviewer without being the spelling every query here filters on.
# A rename leaves every filter matching nothing, so a review that landed reads as none at all.
# A wait then polls out its whole timeout against a review sitting in plain sight.
# `copilot-swe-agent` is the coding agent rather than the reviewer, and does not match this.
READS_AS_REVIEWER = re.compile(r"copilot.*review", re.IGNORECASE)
# The repository this script is hosted in, named because an unrecognized shape is fixed here.
# It is a literal rather than the pull request's own repository.
# That one is where the shape was seen, not where the reader failing on it lives.
HUB = "ptr727/ProjectTemplate"
# The opening tag takes an attribute, as `DETAILS_TAG` already allows on its own pair.
# This reader was the one exception, and the observed format does emit `<details open>`, so a
# `<summary class="x">` read as no summary at all missed a section while reporting no shape.
# An attribute name opens on a letter, which is what tells one from whitespace alone.
# Masking a code span inside `<summary`x`>` otherwise fuses a `<summary >` the body never had.
SUMMARY = re.compile(r"<summary(?:\s+[A-Za-z][^>]*)?>(.*?)</summary>", re.DOTALL | re.IGNORECASE)
TAGS = re.compile(r"</?(?:details|summary)>", re.IGNORECASE)
COUNT = re.compile(r"\((\d+)\)")
# What makes a line a heading rather than prose mentioning the phrase, in either markup.
# A line that is neither still qualifies where it carries a count.
# The section has appeared as a bare line too, and a count is what prose does not carry.
HEADING = re.compile(r"\s*(?:#{1,6}\s|<summary)", re.IGNORECASE)

# How many of the newest reviews and comments both queries read.
# A narrow window drops the reviewer's answer behind ordinary discussion, reporting no answer.
# A test holds this equal to the number the queries carry, since a drift between them reads clean.
WINDOW = 100

# How many of the pull request's own changed files the full query asks for.
# 100 is the connection's own ceiling, so a longer diff is truncated rather than paged.
# A truncated list is reported as one and compared against nothing.
# A path missing from a short window reads exactly like a path the reviewer left out.
# The record holds a pull request of 301 changed files, so the truncation is reachable here.
FILES_WINDOW = 100
# The REST compare endpoint's own ceiling on the file list it returns, which it does not flag.
COMPARE_FILE_CAP = 300
READ_RETRY_DELAYS = (2, 5)
# What a git ref may not hold before it is interpolated into a REST path.
# Refused rather than allow-listed, git permitting more than a branch name usually carries.
# An allow-list of the usual ASCII refused `release/1.0+build` and any non-ASCII name alike.
# Each of those is a ref git accepts, and refusing one refuses the carry on every pull request.
# What is refused instead is the set that changes the path rather than traveling along it.
# `#` truncates the path at a fragment, `?` at a query, and `%` opens an escape.
# A space and a control character are not path characters at all.
UNSAFE_REF = re.compile(r"[\s#?%]|[\x00-\x1f\x7f]")
# A dot segment travels up the path and reaches another repository, which no ref shape excuses.
# Git forbids `..` in a ref name, so this refuses a value that never came from a ref at all.
DOT_SEGMENT = re.compile(r"(?:^|/)\.\.(?:/|$)")

# How many rollup contexts the full query asks for, which is not the window above.
# The query is built from this and the truncation line quotes it, so the two cannot drift.
# Naming it beside a hard-coded literal only documented the literal, which a case then held.
# A case is not the guarantee though, since it holds only where someone runs it.
CHECKS_WINDOW = 100

# How many of the repository's own pull requests, and how much of each one's own reviewer history within them, the repo-wide history read looks across.
# Smaller than WINDOW on purpose, since this is a best-effort signal rather than the certainty WINDOW's own two queries are read for.
# Neither carries a `hasPreviousPage` guard for that same reason.
# A caller finding the signal wrong once quota has plainly reset already has --ignore-quota-signal for it, which is a cheaper remedy than paginating this window out.
HISTORY_PRS = 20
HISTORY_REVIEWS = 20
HISTORY_COMMENTS = 5

# The one-time widened retry `copilot_history` reaches for where HISTORY_PRS carries no Copilot activity at all, rather than reverting every caller to blind polling for the rest of an outage that outlasts it.
# GitHub's own connection ceiling for a single `first`, the same reason FILES_WINDOW and CHECKS_WINDOW hold it.
# Reaching further back needs cursor pagination, which this best-effort signal is not worth paying for, so one wider try is where this stops.
HISTORY_PRS_WIDE = 100

# Liveness query: timestamps and ids only, no comment or review bodies.
# A liveness check does not need the finding text, and re-fetching bodies was 76% of polls.
# It does need the reviewer's non-review answers.
# A wait reading formal reviews alone treats a refusal as an unmet condition.
Q_LIVE = """
query($o:String!,$r:String!,$n:Int!){
  repository(owner:$o,name:$r){ pullRequest(number:$n){
    id headRefOid
    reviews(last:100){ nodes{ author{login} state commit{oid} submittedAt } pageInfo{ hasPreviousPage } }
    comments(last:100){ nodes{ author{login} createdAt } pageInfo{ hasPreviousPage } }
    reviewRequests(first:10){ nodes{ requestedReviewer{ __typename ... on Bot{login} ... on User{login} } } }
  }}}
"""

# The reviewer's own review and comment history, read live from the repository's own pull requests rather than hand-typed or cached across runs.
# Two callers share this one query.
# The bot's node id belongs to the reviewer account rather than to any one PR or review, so it is identical wherever it is found and recency does not matter once one is found.
# A repo-wide reading of the reviewer's own most recent activity is the other caller, and recency is the whole point there, so `submittedAt`, `createdAt`, and `body` ride along for a reader that does not stop at the first bot id.
# Ordered by `UPDATED_AT` rather than `CREATED_AT`, since a fresh review round bumps a pull request's own update time regardless of how long ago it was opened, where creation order can leave a re-reviewed older pull request outside the window entirely.
# `reviews(last:$reviews)` reads the newest rounds a pull request carries rather than the oldest, and `comments(last:$comments)` catches a plain answer that supersedes a formal review without needing every comment a busy pull request holds.
# `$prs` is a variable rather than baked into the document, unlike the other windows in this file, because `copilot_history` runs this same query twice on an empty narrow read, once at HISTORY_PRS and, only then, once at HISTORY_PRS_WIDE.
# One document read at two widths costs nothing a second document would not, and it keeps the two reads provably identical apart from that one number.
# `pullRequests` names no `states` argument, because the connection's omitted default is every state rather than `OPEN`.
# Merged pull requests therefore fill the window alongside open ones.
Q_BOT_ID = """
query($o:String!,$r:String!,$prs:Int!,$reviews:Int!,$comments:Int!){
  repository(owner:$o,name:$r){
    pullRequests(first:$prs, orderBy:{field:UPDATED_AT, direction:DESC}){
      nodes{ number
        reviews(last:$reviews){ nodes{ author{ __typename login ... on Bot{ id } } state body submittedAt } }
        comments(last:$comments){ nodes{ author{ login } body createdAt } }
      } } } }
"""

REQUEST_STATE = """
    reviewRequests(first:100){ nodes{ requestedReviewer{ __typename ... on Bot{login} ... on User{login} } } }
    timelineItems(last:100, itemTypes:[REVIEW_REQUESTED_EVENT]){ nodes{
      ... on ReviewRequestedEvent{ id requestedReviewer{ __typename ... on Bot{login} ... on User{login} } } } }
"""
Q_REQUEST_STATE = """
query($o:String!,$r:String!,$n:Int!){
  repository(owner:$o,name:$r){ pullRequest(number:$n){ id __REQUEST_STATE__ }}}
""".replace("__REQUEST_STATE__", REQUEST_STATE)
M_REQUEST_REVIEWS = """
mutation($pr:ID!,$bot:ID!){
  requestReviews(input:{pullRequestId:$pr, botIds:[$bot], union:true}){ pullRequest{ id __REQUEST_STATE__ } }}
""".replace("__REQUEST_STATE__", REQUEST_STATE)

# The rollup rides this query rather than a REST call, so reading the checks costs no round-trip.
# It is asked of the last commit because a rollup hangs off a commit object.
# A case holds that commit equal to `headRefOid`, since a rollup a push ago still renders whole.
HELD_FIELDS = """
    headRefOid baseRefName state mergeable mergeStateStatus
    baseRepository{ defaultBranchRef{ name } }
    reviews(last:100){ nodes{ id author{login} state commit{oid} submittedAt body } pageInfo{ hasPreviousPage } }
    comments(last:100){ nodes{ author{login} authorAssociation createdAt body } pageInfo{ hasPreviousPage } }
    reviewRequests(first:10){ nodes{ requestedReviewer{ __typename ... on Bot{login} ... on User{login} } } }
    commits(last:1){ nodes{ commit{ oid statusCheckRollup{ state
      contexts(first:__CHECKS_WINDOW__){ pageInfo{ hasNextPage } nodes{
        __typename
        ... on CheckRun{ name status conclusion startedAt isRequired(pullRequestNumber:$n)
          checkSuite{ databaseId app{ slug } workflowRun{ workflow{ databaseId } } } }
        ... on StatusContext{ context state createdAt isRequired(pullRequestNumber:$n) }
      }}}}}}
""".replace("__CHECKS_WINDOW__", str(CHECKS_WINDOW))
Q_HELD = """
query($o:String!,$r:String!,$n:Int!){
  repository(owner:$o,name:$r){ pullRequest(number:$n){ __HELD_FIELDS__ }}}
""".replace("__HELD_FIELDS__", HELD_FIELDS)
Q_FULL = """
query($o:String!,$r:String!,$n:Int!){
  repository(owner:$o,name:$r){ pullRequest(number:$n){ __HELD_FIELDS__
    reviewThreads(first:100){ nodes{ id isResolved
      comments(first:1){ nodes{ author{login} path line body fullDatabaseId pullRequestReview{ id } } } } pageInfo{ hasNextPage } }
    files(first:__FILES_WINDOW__){ pageInfo{ hasNextPage } nodes{ path } }
  }}}
""".replace("__HELD_FIELDS__", HELD_FIELDS).replace("__FILES_WINDOW__", str(FILES_WINDOW))
# Substituted rather than interpolated, because GraphQL is braces from end to end.
# An f-string would need every one of them doubled, which is unreadable against the schema.


# The description and the branch it describes, read together so neither is stale against the other.
# No review or comment connection here, since a description is neither.
Q_CLAIMS = """
query($o:String!,$r:String!,$n:Int!){
  repository(owner:$o,name:$r){ pullRequest(number:$n){ headRefOid body }}}
"""

# A `uses:` reference quoted in a description, in the spelling a workflow writes it.
BODY_USES = re.compile(r'uses:\s*(?P<ref>[A-Za-z0-9_.\-]+/[^\s`"\']+@[^\s`"\']+)')
# A commit the description claims this branch carries, being a verb plus a SHA rather than a SHA.
# The free scan was built first and the corpus rejected it outright.
# Over 25 merged pull requests it raised four findings and every one was correct prose.
# Those were a develop commit named as history, a SHA in quoted output, and two in another repo.
# Nothing in the shape of a bare SHA separates one of those from a genuine claim.
# Separating them by meaning is the similarity heuristic `spec/section-model.md` rejects.
# A path arm measured on the same corpus did worse, flagging 54 of 215 backticked candidates, mostly bare basenames and other repos.
# The verb is what makes a SHA a claim about this branch rather than a mention of one.
# This alternation raises exactly one reference over the same 25, and that one is true.
# The vocabulary is an inclusion list, so a phrasing nobody thought of costs a detection.
# Being incomplete in that direction is the safe one, since it never invents a finding.
BODY_CLAIM = re.compile(
    r"\b(?:fixed|landed|shipped|added|introduced|corrected|resolved|carried|amended)"
    r"\s+(?:in|by|as)\s+`?(?P<sha>[0-9a-f]{7,40})`?",
    re.IGNORECASE,
)
# What `gh` prints when GitHub answered, as opposed to when nothing was reached at all.
HTTP_STATUS = re.compile(r"\(HTTP (\d{3})\)")
# The two GitHub returns for an object that is not there.
# A network error carries no status, and 401 and 403 are credentials and a rate limit.
# Reading one of those as absence reports a correct description as stale.
# Everything outside this set is therefore "not read" rather than "not there".
ABSENT = {"404", "422"}
# What a reference reads as where GitHub never answered for it.
UNREAD = "unread"
# Longer than the ordinary read's, since this one downloads a repository rather than a field.
TARBALL_TIMEOUT = 120

# Threads for the reply path, paginated.
# A first page read as the whole set reports no match on a thread that is simply further along.
# `line` is read for the confirmation line rather than for matching, since a push moves it.
# That is how a reply keyed on a line number went to a thread that had shifted underneath it.
Q_THREADS = """
query($o:String!,$r:String!,$n:Int!,$after:String){
  repository(owner:$o,name:$r){ pullRequest(number:$n){
    reviewThreads(first:100, after:$after){
      nodes{ id isResolved path line comments(first:1){ nodes{ author{login} body } } }
      pageInfo{ hasNextPage endCursor } }
  }}}
"""

# The PR id for a conversation comment is read in the same run that writes it.
Q_COMMENT_TARGET = """
query($o:String!,$r:String!,$n:Int!){
  repository(owner:$o,name:$r){ pullRequest(number:$n){ id url } }}
"""

Q_ATTEST_TARGET = """
query($o:String!,$r:String!,$n:Int!){
  repository(owner:$o,name:$r){ pullRequest(number:$n){ headRefOid baseRefName } }}
"""
ATTESTATION = re.compile(
    r"^<!-- fleet-local-review: head=([0-9a-f]{40}) base=(\S+) findings=(\d+|unknown) -->$",
    re.MULTILINE,
)
TRUSTED_ASSOCIATIONS = frozenset({"OWNER", "MEMBER", "COLLABORATOR"})
LOCAL_REVIEW = Path(__file__).resolve().parent / "local_review.py"

# The conversation-comment and thread mutations the runbook publishes.
# `url` is fetched because it is the one field that confirms a comment or reply carried a body.
# A reply that posted empty still returns a comment, and three did, each then resolved.
M_COMMENT = """
mutation($subjectId:ID!,$body:String!){
  addComment(input:{subjectId:$subjectId, body:$body}){
    commentEdge{ node{ id url body } } }}
"""
M_REPLY = """
mutation($threadId:ID!,$body:String!){
  addPullRequestReviewThreadReply(input:{pullRequestReviewThreadId:$threadId, body:$body}){
    comment{ id url body } }}
"""
M_RESOLVE = """
mutation($threadId:ID!){
  resolveReviewThread(input:{threadId:$threadId}){ thread{ id isResolved } }}
"""


WORKFLOW_KEYS_UNAVAILABLE = "workflowKeysUnavailable"


def gh_graphql(query: str, **variables) -> dict:
    """Run one GraphQL document and return its `data`, raising rather than reporting a blank.

    A string goes through `-f` and an int through `-F`, because `-F` infers a type from the text:
    a reply body of `123` or `true` arrives as an Int or a Boolean and the mutation fails on a
    type nobody passed it, and a body opening with `@` is read as a filename.

    `errors` is checked rather than trusted to the exit code, since a GraphQL document can fail
    per-field while the request itself succeeds, and the caller would read the null that leaves.

    `gh api graphql` exits non-zero with the JSON still on stdout when the response carries errors,
    so that stdout is read first. Only an error set made of `workflowRun` paths is tolerated.
    """
    # Every read below decodes as UTF-8 rather than as whatever the platform's locale is.
    # `gh` emits UTF-8 on every platform, where a Windows console locale is cp1252.
    # A review body carrying one typographic quote crashed the decode and left the caller reading a null stdout.
    argv = ["gh", "api", "graphql", "-f", f"query={query}"]
    for name, value in variables.items():
        argv += ["-F" if isinstance(value, int) else "-f", f"{name}={value}"]
    r = subprocess.run(argv, capture_output=True, text=True, encoding="utf-8", check=False)
    try:
        payload = json.loads(r.stdout)
    except ValueError:
        payload = None
    if not isinstance(payload, dict) or (r.returncode != 0 and not payload.get("errors")):
        sys.stderr.write(r.stderr[:800])
        raise SystemExit(f"gh graphql failed rc={r.returncode}")
    errors = payload.get("errors")
    if errors and not all(_is_workflow_run_error(e) for e in errors):
        sys.stderr.write(json.dumps(errors)[:800])
        raise SystemExit("gh graphql reported errors")
    if payload.get("data") is None:
        sys.stderr.write(json.dumps(errors)[:800])
        raise SystemExit("gh graphql returned no data")
    data = payload["data"]
    if errors:
        sys.stderr.write("gh graphql: workflowRun unreadable, checks are not deduped by workflow\n")
        data[WORKFLOW_KEYS_UNAVAILABLE] = True
    return data


def _is_workflow_run_error(error: object) -> bool:
    """True where the error is a per-field failure on `workflowRun` and nothing else.

    A token without Actions read access can fail that one field, and the field is null in `data`.
    `check_nodes` then keys Actions runs by suite instead, so the digest still reads.
    """
    path = error.get("path") if isinstance(error, dict) else None
    return isinstance(path, list) and bool(path) and path[-1] == "workflowRun"


def gql(query: str, owner: str, repo: str, num: int) -> dict:
    data = gh_graphql(query, o=owner, r=repo, n=num)
    pr = data["repository"]["pullRequest"]
    if data.get(WORKFLOW_KEYS_UNAVAILABLE) and isinstance(pr, dict):
        pr[WORKFLOW_KEYS_UNAVAILABLE] = True
    return pr


def reviewer_requested(pr: dict) -> bool:
    """True where the reviewer sits in the pending request set.

    Read from GraphQL rather than `gh pr view --json reviewRequests`, which omits a Bot
    reviewer entirely and reports an empty set while the reviewer is sitting in it.
    """
    return any(
        (n.get("requestedReviewer") or {}).get("login") == REVIEWER
        for n in ((pr.get("reviewRequests") or {}).get("nodes") or [])
    )


def copilot_history(owner: str, repo: str) -> list[tuple[int, dict]]:
    """The reviewer's own reviews and comments across the repository's most recently updated
    pull requests, newest activity first regardless of which connection it came from.

    Read at HISTORY_PRS first and, only where that comes back carrying no usable bot id, read
    again at the wider HISTORY_PRS_WIDE. A narrow window emptying out is the ordinary case, a
    repository whose most recent activity genuinely carries none of the reviewer's, and costs
    nothing beyond the one call either caller below was always going to make. It stops being
    ordinary once an outage outlasts HISTORY_PRS pull requests: every one of them then carries
    the same silence, the narrow window empties out too, and both callers would otherwise fall
    back to blind polling for the rest of the outage with no way to tell that outage apart from a
    repository that has simply never seen a Copilot review, a shape reproduced over four
    consecutive pull requests. The wider read is what tells the two apart.

    Emptying out is not the only way the narrow window fails a bot-id lookup, though: it can
    carry real activity and still have none, when every entry within it is a plain comment. A
    formal review, `copilot_bot_id`'s only source for the id, can sit just outside the narrow
    window while a newer comment sits inside it, and returning the narrow read as soon as it has
    anything at all left that review permanently unread. Widening is keyed on
    `copilot_bot_id(entries)` rather than on emptiness for exactly that case, so a comment-only
    narrow window still triggers the wider read the same way an empty one does. The ordinary case
    -- a narrow window already carrying a review -- still costs one call rather than two, since
    that is the common shape a usable bot id already satisfies.
    """
    entries = _copilot_history_window(owner, repo, HISTORY_PRS)
    if entries and copilot_bot_id(entries) is not None:
        return entries
    return _copilot_history_window(owner, repo, HISTORY_PRS_WIDE)


def _copilot_history_window(owner: str, repo: str, prs: int) -> list[tuple[int, dict]]:
    """One read of `copilot_history`'s traversal, over the `prs` most recently updated pull
    requests. Split out so the narrow and widened reads share every line but the one count.

    Paired with the pull request number each came from, since a repo-wide reading has to name
    where it looked rather than just assert one. Sorted by timestamp rather than trusted to
    arrive in that order, since the two connections are read separately and merged.

    Unfiltered, including the caller's own pull request if it appears in the window. No pull
    request is special-cased out: `copilot_bot_id` is happy to read a valid id from anywhere,
    including an earlier round on the very pull request `wait` is running against, and
    `quota_signal` treats an earlier round on that same pull request as real evidence about the
    account rather than a self-reference to discard. A refusal on the caller's own current head
    is caught directly, at higher priority, before either reading here is ever consulted.

    A plain comment is read alongside a formal review, tagged `_kind` so `copilot_bot_id` can
    still tell them apart, because `answered_outside_review` already treats a comment as
    meaningful reviewer activity for one pull request, and a repo-wide reading that only checked
    reviews could keep reporting a stale refusal after a newer plain comment answered it.

    One call serves both the bot-id lookup and the repo-wide quota reading below, since both
    need the same traversal and a caller needing either otherwise pays for it twice.
    """
    data = gh_graphql(
        Q_BOT_ID, o=owner, r=repo, prs=prs, reviews=HISTORY_REVIEWS, comments=HISTORY_COMMENTS
    )
    nodes = ((data.get("repository") or {}).get("pullRequests") or {}).get("nodes") or []
    entries = []
    for node in nodes:
        number = node.get("number")
        for review in (node.get("reviews") or {}).get("nodes") or []:
            author = review.get("author") or {}
            if author.get("__typename") == "Bot" and author.get("login") == REVIEWER:
                entries.append(
                    (number, {**review, "_at": review.get("submittedAt") or "", "_kind": "review"})
                )
        for comment in (node.get("comments") or {}).get("nodes") or []:
            if (comment.get("author") or {}).get("login") == REVIEWER:
                entries.append(
                    (number, {**comment, "_at": comment.get("createdAt") or "", "_kind": "comment"})
                )
    return sorted(entries, key=lambda pair: pair[1]["_at"], reverse=True)


def copilot_bot_id(history: list[tuple[int, dict]]) -> str | None:
    """The Copilot reviewer bot's node id from an already-fetched history, or None where none
    carries one.

    Read from the review entries only, since a comment's author carries no node id in the query
    that fetches it, that field belonging to the review connection's `... on Bot{ id }` selection.
    None is a real answer, not a failure to retry: a repository's very first PR, or one Copilot
    has never reviewed, carries no review to read the id from. The caller falls back to polling
    only rather than guessing, since a fabricated or reused id resolves globally and a wrong one
    would silently target another repository's PR.
    """
    for _number, entry in history:
        if entry.get("_kind") != "review":
            continue
        bot_id = (entry.get("author") or {}).get("id")
        if bot_id:
            return bot_id
    return None


def quota_signal(history: list[tuple[int, dict]]) -> tuple[int, dict] | None:
    """The reviewer's own most recent activity anywhere in this repository, review or comment,
    where it is a quota refusal, paired with the pull request number it came from.

    The most recent rather than any match, since a repository can carry an old refusal and a
    later working round both, and only the most recent record says which is still true. A
    genuine review anywhere since the refusal spends it, the same reading `refusing_review`
    already gives one pull request's own history, and so does a later plain comment, since
    `answered_outside_review` treats one the same way for that one pull request. Reading across
    pull requests rather than one is what a repository-wide account state needs, since the pull
    request in front of a caller can carry none of its own reviewer activity to read that state
    from at all.
    """
    newest = history[0] if history else None
    return newest if newest and quota_refusal(newest[1]) else None


def rate_limited_by(pr: dict, login: str) -> str | None:
    """The service name a rate-limit marker gives, where this login's newest comment or review
    on this pull request carries one. `None` otherwise.

    Reads both connections since the one observed marker, CodeRabbit's, rides a
    plain PR comment rather than a formal review, and a future bot using the same convention
    might use either. Current pull request only, unlike Copilot's `quota_signal`: that one
    generalizes from a pattern of silence across several pull requests, and this has one
    observed instance to generalize from, not a history to read one out of.
    """
    nodes = reviewer_nodes(pr, "comments", login) + reviewer_nodes(pr, "reviews", login)
    if not nodes:
        return None
    newest = max(nodes, key=lambda n: n.get("createdAt") or n.get("submittedAt") or "")
    match = RATE_LIMITED.search(newest.get("body") or "")
    return match.group(1) if match else None


def newest_request_event(pr: dict) -> str | None:
    """The id of the newest review-request event naming the reviewer, or None where there is none.

    The id is compared rather than a count, since a count over a bounded window stops moving once
    the window is full.
    """
    ids = [
        n.get("id")
        for n in ((pr.get("timelineItems") or {}).get("nodes") or [])
        if (n.get("requestedReviewer") or {}).get("login") == REVIEWER
    ]
    return ids[-1] if ids else None


def request_recorded(before: dict, after: dict) -> bool | None:
    """Whether a review request that returned success left anything on the pull request.

    True where the reviewer sits in the pending set afterwards, or where a review-request event
    naming it is newer than the newest one read before the request. False where neither holds,
    which is the shape an exhausted Copilot allowance left: every mutation succeeded, and none
    added an event or a pending reviewer. None where the answer lacks either field, since an
    unread state is not a reading of one, and the wait then polls as it did before this existed.
    None too where a login that reads as the reviewer's but is spelled otherwise answers, since
    a drifted login would otherwise read a recorded request as an unrecorded one and blame the
    account for it.
    """
    if after.get("reviewRequests") is None or after.get("timelineItems") is None:
        return None
    nodes = after["reviewRequests"].get("nodes") or []
    nodes = nodes + (after["timelineItems"].get("nodes") or [])
    for n in nodes:
        login = (n.get("requestedReviewer") or {}).get("login") or ""
        if login != REVIEWER and READS_AS_REVIEWER.search(login):
            return None
    if reviewer_requested(after):
        return True
    return newest_request_event(after) != newest_request_event(before)


def request_copilot_review(
    owner: str, repo: str, num: int, pr_node_id: str, bot_id: str | None, settle: float
) -> tuple[str, bool | None]:
    """Ask Copilot to review the current head, and say in one line what happened.

    This exists because `wait` used to only ever poll, never request, so a PR whose auto-seed
    never fired or whose prior request was superseded by a later push sat waiting the full
    timeout for a review nothing had asked for, twice in one session before this was written.
    `union:true` adds the bot to the request set rather than replacing it, so a human reviewer
    requested alongside it stays requested. Raises where `gh_graphql` does, on the mutation
    itself, the same as every other write in this script: a failed write is reported, never
    quietly swallowed into a fallback that would recreate the exact blind-poll failure this
    function exists to prevent, from a different cause. The one quiet path is finding no bot id
    to request with at all, which is not a failure, since a repository with no Copilot review
    anywhere carries nothing to read the id from. The id itself is the caller's to find, via
    `copilot_bot_id` over a `copilot_history` read it already paid for, which already tried both
    the narrow HISTORY_PRS window and the wider HISTORY_PRS_WIDE one before coming up empty.

    The second value is `request_recorded` over the state read before the request and the one
    the mutation answers with, and None where nothing was requested. An answer reading as
    unrecorded is read once more with the same query and judged the same way, `settle` seconds
    later, since GitHub can take a moment to record a request it has already accepted.
    """
    if not bot_id:
        return (
            (
                f"no Copilot review found across the last {HISTORY_PRS} or, widened once for "
                f"exactly this reason, the last {HISTORY_PRS_WIDE} most-recently-updated pull "
                "requests to read the reviewer bot id from, so nothing was requested here, "
                "polling only. Seed one via the UI if this repository has never had one at all."
            ),
            None,
        )
    before = gh_graphql(Q_REQUEST_STATE, o=owner, r=repo, n=num)["repository"]["pullRequest"]
    answer = gh_graphql(M_REQUEST_REVIEWS, pr=pr_node_id, bot=bot_id)
    after = (answer.get("requestReviews") or {}).get("pullRequest") or {}
    state = backoff(
        after,
        lambda: gh_graphql(Q_REQUEST_STATE, o=owner, r=repo, n=num)["repository"]["pullRequest"],
        lambda p: request_recorded(before, p or {}) is False,
        delays=(settle,),
    )
    recorded = request_recorded(before, state or {})
    line = f"requested a Copilot review on the current head (bot {bot_id})"
    return line, recorded


def reviewer_nodes(pr: dict, field: str, login: str = REVIEWER) -> list[dict]:
    """This login's own nodes under `field`, oldest first as the API returns them.

    `login` defaults to the fully-modeled reviewer, so every existing call keeps reading
    Copilot's own nodes unchanged. A caller reading another known reviewer's presence
    generically, identity and commit only, no body parsing, passes it explicitly.
    """
    return [
        n
        for n in ((pr.get(field) or {}).get("nodes") or [])
        if (n.get("author") or {}).get("login") == login
    ]


def answered_outside_review(pr: dict) -> dict | None:
    """The reviewer's newest plain comment, where it postdates its newest formal review.

    The test is the shape of the answer rather than its cause, which this reads nothing of:
    a comment carries no commit, so it satisfies no coverage check whatever it says.
    Treating that shape as an unmet condition is what leaves a wait with nothing at its end.
    A comment older than the newest review is spent, since the review it preceded did land.
    """
    comments = reviewer_nodes(pr, "comments")
    # A blind review window leaves no honest baseline to date a comment against.
    # An empty one dates every comment as newer, so each reads as an answer.
    # Reporting nothing keeps the wait polling, where a wrong answer ends it outright.
    if not comments or window_blind(pr, "reviews"):
        return None
    newest = max(comments, key=lambda n: n.get("createdAt") or "")
    reviews = reviewer_nodes(pr, "reviews")
    latest_review = max((n.get("submittedAt") or "" for n in reviews), default="")
    return newest if (newest.get("createdAt") or "") > latest_review else None


def window_blind(pr: dict, field: str, login: str = REVIEWER) -> bool:
    """True where this login's own nodes can sit behind the window, so the view cannot decide.

    `login` defaults to the fully-modeled reviewer, so every existing caller keeps reading
    Copilot's own blind spot unchanged. A caller reading another known reviewer's own window,
    Qodo's `comments` connection carrying its findings comment, passes it explicitly.

    Each query reads the newest nodes rather than this login's, so ordinary traffic is what
    pushes theirs out of reach. Nodes arrive in creation order, so anything behind the window is
    older than everything inside it: one of this login's in view bounds every hidden one as
    older still, which settles the question rather than leaving it open. `hasPreviousPage` is
    what says anything is back there at all, since a full window and a window holding the lot
    are the same length.
    """
    older = ((pr.get(field) or {}).get("pageInfo") or {}).get("hasPreviousPage")
    return bool(older) and not reviewer_nodes(pr, field, login)


def threads_truncated(pr: dict) -> bool:
    """True where `reviewThreads(first:100)` cut off before this pull request's actual thread
    count, so `threads=`/`unresolved=` below undercount rather than reading the true total.

    Unlike `window_blind`'s `last`-windowed connections, `reviewThreads` reads forward from the
    first page in the connection's own creation order, so `hasNextPage` here means the threads
    left unread are the newest ones, not the oldest, exactly the ones most likely to still be
    open. That inversion is why this is its own guard rather than a second use of `window_blind`:
    that one settles the question from what is already in view, and there is no such settling
    available here, only the fact that something was cut, undercounted ever since `unresolved`
    widened from Copilot's own threads to every open thread.
    """
    return bool(((pr.get("reviewThreads") or {}).get("pageInfo") or {}).get("hasNextPage"))


def reviews_truncated(pr: dict) -> bool:
    """True where the `reviews` connection's own page cut off before this pull request's actual
    round count, so a suppressed or outside-diff-range finding raised on an older, windowed-out
    round is missing from `suppressed=`/`cr_outside_diff=` below rather than merely counted as
    stale.

    Unlike `window_blind`, which settles a coverage or dating question once even one of a
    login's own reviews sits in view, a finding count needs every round read, not just a
    recent-enough one: an unread round can carry a finding of its own regardless of whether a
    newer round from the same login is visible. `hasPreviousPage` on this `last`-windowed
    connection means the reviews left unread are the oldest, exactly where an unanswered
    finding is most likely to still sit unresolved.
    """
    return bool(((pr.get("reviews") or {}).get("pageInfo") or {}).get("hasPreviousPage"))


def refusal_of(node: dict) -> str:
    """The review's body where its opening line says the reviewer did not review, otherwise empty.

    Read over the opening rather than the whole body, because a refusal replaces the review and
    is the only thing the body carries, where a review that merely quotes the wording carries it
    below its own overview. This script and its documentation are that quotation, so a body-wide
    match would report the pull request adding this check as a refusal, which is the false
    positive the suppressed-block matcher already had once.

    The opening is one line rather than two, and the second line is where the cost of widening
    it shows: a review's first line is its heading and its second is the overview prose, which
    is exactly where a review describing this check states the wording. Reading two lines passed
    every case here except that one, which is the case that matters. A refusal introduced by a
    heading would sit below the opening and be missed, and answering that shape means telling it
    from an overview rather than reading one line further.

    The second overview format's version marker is skipped rather than read as that opening,
    since it is markup that renders invisibly and a refusal wearing it opens on the line below.
    That leaves the heading-introduced refusal named above missing in both formats alike rather
    than in one, which is where it was before this format arrived. Such a round states no
    coverage either, so it reaches the maintainer as an unproven diff rather than as a clean
    pass, which is a worse remedy than a refusal's own and not a false green.
    """
    body = node.get("body") or ""
    opening = next(
        (ln for ln in body.splitlines() if ln.strip() and not CCR_OVERVIEW.fullmatch(ln)), ""
    )
    return body if REFUSAL.search(opening) else ""


def quota_refusal(node: dict) -> bool:
    """True where a refusal on this node names the account quota rather than another cause.

    Read from `refusal_of`'s own return rather than the raw body, so its exemptions protect
    this reading too: a node that is not a refusal at all returns `""` here, on which `QUOTA`
    cannot match, rather than this needing its own opening-line anchor and fenced-quote guard.
    The distinction matters because the two remedies are opposite. A file-count refusal is
    cleared by splitting the pull request. A quota one is an account-level state that a
    re-request or a wait does not touch, so treating the two alike sends a reader at a remedy
    that does nothing here.
    """
    return bool(QUOTA.search(refusal_of(node)))


def possible_quota(node: dict) -> bool:
    """True where a refusal on this node names an error rather than any cause.

    The weekly rate limit posts exactly this body, its cause written only to the job log, so it
    reads as a possible quota hit rather than as a transient failure. A re-request into a reached
    limit spends quota it cannot recover and returns the same body, which is why one is enough.
    """
    return bool(ERROR_REFUSAL.search(refusal_of(node)))


@functools.cache
def run_cause(owner: str, repo: str, oid: str, before: str) -> tuple[str, str] | None:
    """The error type and rate-limit sentence the reviewer's own failed run logged, or None.

    An error refusal's body names no cause, and the run that produced it does: the reviewer runs
    as an Actions workflow on the reviewed commit, and its job log states the weekly limit and
    when it resets. The run read is the newest failed one on that commit created no later than
    the review, so a later run on the same commit is not read as this round's.

    None wherever any read fails or finds nothing, which leaves the round a possible quota hit
    rather than a confirmed one or a cleared one. Memoized for the run, so the digest and the exit
    code read one log once.
    """
    if not oid or not ISO_STAMP.fullmatch(before):
        return None
    proc = gh_rest(
        f"repos/{owner}/{repo}/actions/runs?event=dynamic&head_sha={oid}&per_page=50",
        f'[.workflow_runs[] | select(.path == "{COPILOT_RUN_PATH}" and .conclusion == "failure"'
        f' and .created_at <= "{before}")] | max_by(.created_at) | .id // empty',
    )
    run = proc.stdout.strip()
    if proc.returncode != 0 or not run.isdigit():
        return None
    proc = gh_rest(f"repos/{owner}/{repo}/actions/runs/{run}/jobs", ".jobs[0].id // empty")
    job = proc.stdout.strip()
    if proc.returncode != 0 or not job.isdigit():
        return None
    proc = gh_rest(f"repos/{owner}/{repo}/actions/jobs/{job}/logs", raw=True)
    if proc.returncode != 0:
        return None
    log = ANSI.sub("", proc.stdout)
    kinds, said = RUN_ERROR_TYPE.findall(log), RUN_RATE_LIMIT.search(log)
    if not kinds:
        return None
    kind = "rate_limit" if "rate_limit" in kinds else kinds[-1]
    return kind, said.group(1).strip() if said else ""


def confirmed_quota(owner: str, repo: str, node: dict) -> str:
    """The run log's rate-limit sentence where an error refusal's run logged a rate limit, else ""."""
    if not possible_quota(node):
        return ""
    cause = run_cause(
        owner, repo, (node.get("commit") or {}).get("oid") or "", node.get("submittedAt") or ""
    )
    if cause is None or cause[0] != "rate_limit":
        return ""
    return cause[1] or "the run logged a rate limit and stated no reset time"


def stopping_refusal(pr: dict) -> dict | None:
    """The reviewer's newest review on this pull request, on any head, where it is a quota refusal or an error.

    Read across heads rather than on the head alone, since a push after such a refusal moves the
    head and the account state the refusal reports does not move with it. A genuine round since
    spends it, the newest review being the one read, and so does a plain comment of the
    reviewer's since, which `answered_outside_review` already reads as spending one.
    """
    if answered_outside_review(pr):
        return None
    newest = newest_of(reviewer_nodes(pr, "reviews"))
    if newest is None or not (quota_refusal(newest) or possible_quota(newest)):
        return None
    return newest


def refusing_review(pr: dict) -> dict | None:
    """The reviewer's newest refusal carrying the current head, where one is there.

    Head-scoped, unlike a suppressed finding, because a refusal is a statement about one commit:
    a push retires it, and the round the push raises either reviews that head or refuses it in
    its own right.

    A refusal alongside a genuine review of the same head is spent too, and that is the caller's
    reading rather than this one's, since what spends it is coverage this cannot see from a
    refusal alone. Both callers hold it: `main` returns 0 on `reviewed_head` before reaching
    here, and the digest reports the field only where nothing covers the head.
    """
    head = pr["headRefOid"]
    refusals = [
        n
        for n in reviewer_nodes(pr, "reviews")
        if (n.get("commit") or {}).get("oid") == head and refusal_of(n)
    ]
    return newest_of(refusals)


def newest_of(nodes: list[dict]) -> dict | None:
    """The newest of these nodes by submission time, the later arrival winning a tie.

    `max` returns the first maximal element and `reviewer_nodes` returns these oldest first, so
    a tie on `submittedAt` selected the oldest round rather than the newest. GitHub stamps two
    rounds on one commit with the same second often enough for that to decide a digest, and the
    position in the connection is the only later-arriving signal there is once the timestamps
    agree.

    A node carrying no timestamp sorts oldest, a round that has not been submitted not being the
    one describing the head.

    Shared rather than spelled per reader, because `effort=` and `overview=` are read from a
    round each and a digest line whose fields disagree about which round it describes is worse
    than either field being wrong on its own.
    """
    if not nodes:
        return None
    return max(enumerate(nodes), key=lambda pair: (pair[1].get("submittedAt") or "", pair[0]))[1]


def head_reviews(pr: dict) -> list[dict]:
    """The reviewer's own reviews that cover the current head, refusals excluded.

    A refusal is not coverage. It is a formal review, `state: COMMENTED`, carrying the head's
    commit and raising no threads, so it satisfies every check a clean pass does and renders a
    digest identical to one. That is how a pull request of 301 changed files, one over the
    reviewer's limit, sat one command from merging on a review that never ran.

    One list rather than a predicate beside a filter, because coverage and the count of rounds on
    the head are read from the same set and a second spelling of it drifts from the first.
    """
    head = pr["headRefOid"]
    return [
        n
        for n in reviewer_nodes(pr, "reviews")
        if (n.get("commit") or {}).get("oid") == head and not refusal_of(n)
    ]


def reviewed_head(pr: dict) -> bool:
    """True where one of the reviewer's own reviews covers the current head's commit.

    The liveness query carries no bodies, so a refusal reads there as ordinary coverage. That is
    deliberate rather than a gap: it ends the wait, which is what a terminal outcome should do,
    and the full read every wait finishes with is what tells the two apart. No exit code and no
    merge decision is taken from the liveness reading.
    """
    return bool(head_reviews(pr))


def head_review_done(pr: dict, min_rounds: int) -> bool:
    """True once a round past position `min_rounds` in the reviewer's own history covers the head.

    Positions are read off `reviewer_nodes(pr, "reviews")`, every one of the reviewer's own
    reviews on this pull request, on any head, refusals included, oldest first: the exact list
    `status` counts for its own `rounds=`. A caller that already read that count, then
    re-requested a review on an already-covered head, passes it through `wait --min-rounds`
    instead, so this holds out for a round sitting past it.

    The round satisfying `min_rounds` and the round satisfying coverage must be the same round,
    which is why this walks one list rather than checking `reviewed_head(pr)` and a count as two
    independent clauses. Two clauses would let a refusal past the baseline satisfy the count
    while an untouched, older genuine round still satisfies `reviewed_head`, reporting done on a
    round that answered the re-request by declining it. A caller passing 0, every caller that
    predates this flag among them, reduces to `reviewed_head` exactly, since position 0 is every
    position a non-empty list has.

    Counting `reviewer_nodes(pr, "reviews")` rather than `head_reviews(pr)`, the head-scoped,
    refusal-excluded list `reviewed_head` itself narrows to, matters independently of the above:
    a push during the wait moves the head, which empties a head-scoped list and reads a brand
    new head's first genuine round as still short of a baseline measured against a head that no
    longer exists, where this list carries forward since a push adds to it rather than resetting
    it. Positions still land past `min_rounds` correctly once the head moves, since a review on
    the superseded head sits at an earlier position than one on the new head, never a later one.

    The same `last:100` window both `Q_LIVE` and `Q_FULL` place on their own `reviews`
    connection bounds this the way it bounds `status`'s own `rounds=`: a pull request whose
    Copilot rounds have aged out of that window under other reviewers' own traffic undercounts
    here exactly as `status` would, reading a landed round as still short of the baseline rather
    than crediting it, so this fails toward a caller waiting longer, not toward a false success.
    """
    revs = reviewer_nodes(pr, "reviews")
    head = pr["headRefOid"]
    return any(
        i >= min_rounds and (n.get("commit") or {}).get("oid") == head and not refusal_of(n)
        for i, n in enumerate(revs)
    )


def review_effort(pr: dict) -> tuple[str, str]:
    """The newest head review's effective effort and selection source.

    GitHub can render an inherited choice as `Default (Lite)`, `Default (Balanced)`, or
    `Default (Max)`. A bare level is explicit. The setting remains user-controlled, and this
    reader only reports metadata that the completed review body exposes.
    """
    newest = newest_of(head_reviews(pr))
    if newest is None:
        return "unknown", "unknown"
    plain = strip_fences(newest.get("body") or "")
    for line in plain.splitlines():
        match = EFFORT_LINE.fullmatch(line)
        if match:
            source = "default" if match.group(1) else "explicit"
            return (match.group(2) or match.group(3)).lower(), source
    return "unknown", "unknown"


def code_indented(line: str) -> bool:
    """Whether the line opens four or more columns in, the indentation of a Markdown code block.

    Measured in columns rather than characters, a tab being four of them and so an indented code
    block on its own, which a character count reads as one column. Only spaces and tabs indent,
    so a line opening on any other whitespace is ordinary text.
    """
    expanded = line.expandtabs(4)
    return len(expanded) - len(expanded.lstrip(" ")) >= 4


def is_coverage_line(line: str) -> bool:
    """Whether this line is the reviewer stating its file coverage, rather than prose about it."""
    if FLEET_REVIEW.search(line):
        return True
    if COVERAGE_BULLET.match(line):
        return True
    return bool(COVERAGE_SENTENCE.match(line)) and "changed file" in line.lower()


def coverage_statements(body: str) -> list[str]:
    """The lines this round states its file coverage on, quotations excluded.

    An inline code span is a quotation the same as a fenced block, so a span carrying a coverage
    line reads as one this round stated rather than one it quoted.

    A blockquoted line is the third such convention and is dropped whole rather than unwrapped.
    The line anchor used to exclude one for free, the marker reader having required the marker to
    be the whole of its line and a `>` not being whitespace. Reading the marker within its line
    took that exclusion away, so the quotation it was refusing is refused here instead, where the
    other two already are. The prose readers beside it are unaffected either way, each being
    anchored to a line opener that a `>` already fails.

    The quotation runs to the blank line that closes it rather than to the end of the line opening
    it, which is Markdown's own lazy continuation: a line carrying no `>` under one that does is
    still rendered inside the quotation. Refusing only the marked line reads the line after a
    quotation's first as this round's own statement.

    Continuation carries paragraph text and nothing else, so a line opening a block of its own ends
    the quotation above it. Of the three spellings read here only the reviewer's own sentence is
    paragraph text, the bullet and the marker each opening a block, and treating those as quoted
    reported no coverage at all over a round that stated it in full.

    A line indented four or more columns is the fourth convention, an indented code block, and a
    marker on one is dropped from the line rather than the line being dropped whole. The indent
    opens a code block only where nothing runs on into it, since an indented line under paragraph
    text is that paragraph's lazy continuation and renders as part of it, and a list item's own
    text runs on the same way. A heading, a setext underline, a thematic break, and a line opening
    an HTML comment each end their own block, and a code line carries its block on to the next. `CCR_OVERVIEW`
    and `CCR_FINDINGS` bound their openers to three spaces for the same reason, and the marker,
    being read within its line, takes that bound here instead of in its pattern. The rest of the
    line is still read, the bullet and the sentence being left to their own readers as before, and
    the marker comes off before the line is kept, because every reader downstream reads the
    kept line with its indentation already stripped.
    """
    plain = CODE_SPAN.sub(" ", strip_fences(body or "", to_end=True))
    found = []
    quoted = False
    continues = False
    for ln in plain.splitlines():
        stripped = ln.lstrip(" \t")
        if not stripped:
            # A blank line is what ends a blockquote, so the next line starts outside one again.
            quoted = False
            continues = False
            continue
        indented = code_indented(ln)
        code = indented and not continues
        continues = not code and not re.match(
            r"#{1,6}(?:\s|$)|([-*_])(?: *\1){2,} *$|(?:=+|-+) *$|<!--", stripped
        )
        # An indented line cannot open a blockquote, being code or a paragraph's continuation, so its `>` is text rather than a quotation's own marker.
        # Read as one, a prompt quoted in a code block opened a blockquote that swallowed every line to the next blank one, and a coverage statement among them read as none stated at all.
        if not indented and BLOCKQUOTE.match(stripped):
            quoted = True
            continue
        if not indented and STARTS_BLOCK.match(stripped):
            # A line opening its own block is not continuation text, so the quotation ends above it.
            quoted = False
        if code:
            ln = FLEET_REVIEW.sub(" ", ln)
        # What is left under a quotation is paragraph text, which is still inside it by Markdown's own lazy continuation and renders as part of it.
        if quoted or not is_coverage_line(ln):
            continue
        found.append(ln.strip())
    return found


def read_coverage(line: str) -> tuple[int, int] | None:
    """The (reviewed, changed) counts the line states, or None where they cannot be believed.

    None covers a wording this has no vetted spelling for and a pair that cannot both be true
    alike, since each leaves the same question unanswered and each takes the same remedy. A round
    reporting it read more files than the pull request changed is not a round that read them all,
    it is a line this script is parsing wrongly, and reading it as full coverage fails open on
    exactly the statement that says something is off.

    Two readings on one line that disagree are that same case, reached within a line rather than
    between two. Reading the marker within its line rather than as the whole of one is what makes a
    line able to carry both, and settling such a line on the marker alone would clear the gate for a
    round whose own prose says it read part of the diff.
    """
    marker = FLEET_REVIEW.search(line)
    m = COVERAGE_COUNTS.search(line)
    prose = None
    if m:
        got = (m.group(1), m.group(2)) if m.group(1) else (m.group(3), m.group(4))
        # Bounded exactly as the marker's own runs are below, and for the same reason.
        # Guarding one of the two left the other crashing the digest on the sentence spelling.
        read = (counted(got[0]), counted(got[1]))
        if read[0] is None or read[1] is None:
            return None
        prose = (read[0], read[1])
    if marker:
        # Judged before conversion, `int` raising above 4300 digits and no review stating a count of even five.
        # A run this long is a line to refuse rather than one to crash on.
        pair = (counted(marker.group(1)), counted(marker.group(2)))
        if pair[0] is None or pair[1] is None:
            return None
        counts = (pair[0], pair[1])
        # A line carrying both readings and disagreeing is a line this cannot believe, which is the impossible pair below reached a different way, and it takes that case's answer.
        # Taking the worse of the two instead reads an impossible pair as the better one, a negative shortfall never being the larger.
        # `reviewed=4 changed=3` beside a prose `4/4` then passed the gate outright, where refusing the line blocks it.
        # One line and two do not settle under the same code, reading unvetted and partial, and neither passes, which is the property that matters here.
        if prose is not None and prose != counts:
            return None
        return counts if counts[0] <= counts[1] else None
    if prose is None:
        return None
    return prose if prose[0] <= prose[1] else None


def coverage_of(node: dict) -> tuple[str, str]:
    """This round's coverage reading, with the line it was read from.

    A round making no statement at all reads as unstated rather than unvetted. 28 of the 333
    bodies measured carry an overview and a change list and nothing more. The shape is recognized,
    but it cannot prove full diff coverage and therefore blocks the status gate.

    A wording that is coverage-shaped and parses to no counts is the failure whose remedy is
    fixing this script, since a gate that allows whatever it does not recognize stops gating as
    the wording drifts, which it has done once already for each of the two patterns above.
    """
    worst, detail = UNSTATED, ""
    for line in coverage_statements(node.get("body") or ""):
        counts = read_coverage(line)
        state = UNVETTED if counts is None else (FULL if counts[0] == counts[1] else PARTIAL)
        if SEVERITY.index(state) < SEVERITY.index(worst):
            worst, detail = state, line
    return worst, detail


def head_coverage(pr: dict) -> tuple[str, str]:
    """The worst coverage the rounds covering the current head state, and the line saying it.

    Head-scoped for the reason a refusal is, and unlike a suppressed finding: a partial round is
    a statement about one commit's diff, and the push that changes that diff raises a round which
    reads the whole of the new one. A refusal needs no exemption of its own here, `head_reviews`
    having dropped it already, and reading one would report the round that declined as a wording
    this script fails to recognize.
    """
    worst, detail = UNSTATED, ""
    for node in head_reviews(pr):
        state, line = coverage_of(node)
        if SEVERITY.index(state) < SEVERITY.index(worst):
            worst, detail = state, line
    return worst, detail


def carried_coverage(pr: dict) -> tuple[str, str, str] | None:
    """The newest round on this pull request that states any coverage, and the commit it read.

    None where no round ever stated any, which is the one case left with nothing to carry.

    A round covering the head and stating no coverage is the ordinary shape of this format rather
    than a defect in it: five of ten rounds measured omit the marker, and both omissions measured
    on one drive were re-reviews of a delta of one line and of four files. Blocking every one of
    those asks for a re-request that produced the marker in one of the four measured, so the
    block clears by chance rather than by asking and a reader learns to route around it.

    What an earlier round's statement is worth depends on what changed since, which is why the
    caller reports that delta beside this rather than this returning a verdict. The state carries
    as it was: a round that read part of the diff carries partial, since the part it never read is
    still unread.

    `newest_of` rather than plain `max`, for the reason it documents: two rounds share a
    timestamp often enough to decide a digest, `max` returns the first maximal element, and
    `reviewer_nodes` returns these oldest first, so a tie selected the older of the two. Returned
    that way a partial round outranked the full one stamped in the same second, or the reverse,
    on nothing but the order the connection happened to list them in.
    """
    stating = {}
    for node in reviewer_nodes(pr, "reviews"):
        if refusal_of(node):
            continue
        state, line = coverage_of(node)
        if state != UNSTATED:
            stating[id(node)] = (node, state, line)
    newest = newest_of([n for n, _, _ in stating.values()])
    if newest is None:
        return None
    _node, state, line = stating[id(newest)]
    return state, line, (newest.get("commit") or {}).get("oid") or ""


def delta_since(owner: str, repo: str, base: str, head: str) -> str:
    """How much changed between the round that stated coverage and the head, as one clause.

    Every reading is a clause its callers complete a sentence with, so a delta that could not be
    read composes as readably as one that could. Returned as a noun phrase instead, the failure
    read "with the delta between them could not be read".

    Reported rather than judged. Whether a delta is small enough to let an earlier round's
    coverage stand is the maintainer's reading, and this gives them the number to read it from.

    An unreadable compare says so rather than returning a zero, since a delta that could not be
    read and a delta of nothing take opposite decisions. The filter defaults neither field for
    that reason: a 200 missing either one fails the filter and reads as unread, where a `//`
    default rendered it as a delta of nothing. The API caps its file list at 300, so a larger
    delta is reported as at least that, never as exactly it.
    """
    if not base or not head:
        return "the commit either side of the delta is unknown"
    if base == head:
        return "nothing changed between them"
    proc = gh_rest(
        f"repos/{owner}/{repo}/compare/{base}...{head}",
        r'if (.files | type) == "array" and (.ahead_by | type) == "number" '
        r'then "\(.files | length) \(.ahead_by)" else empty end',
    )
    if proc.returncode != 0 or not proc.stdout.strip():
        return "the delta between them could not be read"
    try:
        files, commits = (int(x) for x in proc.stdout.split())
    except ValueError:
        return "the delta between them could not be read"
    at_least = "at least " if files >= COMPARE_FILE_CAP else ""
    return (
        f"{at_least}{files} file{'' if files == 1 else 's'} changed over "
        f"{commits} commit{'' if commits == 1 else 's'} since"
    )


def files_shortfall(pr: dict) -> str:
    """Why the payload's changed-file list cannot be matched against a table, or "" where it can.

    One predicate for the table reading's reason and for `files_unread`'s re-read, so the two
    cannot disagree about which lists were read. A list the query cut short is not a shortfall
    here, since that list was read, and `table_shortfall` names it separately.
    """
    files = pr.get("files")
    if not isinstance(files, dict):
        return "the changed-file list is absent from the query, so the table has nothing to match"
    nodes = files.get("nodes")
    page = files.get("pageInfo")
    if (
        not isinstance(nodes, list)
        or not isinstance(page, dict)
        or not isinstance(page.get("hasNextPage"), bool)
        or any(not (n or {}).get("path") for n in nodes)
    ):
        return "the changed-file list is malformed, so the table cannot be matched against it"
    if not nodes and not page["hasNextPage"]:
        return "the pull request changes no files, so the table has nothing to match"
    return ""


def files_unread(pr: dict) -> bool:
    """Whether the changed-file list is the reason a file table cannot stand in for this head.

    The table reading runs only where a round covers the head and states no coverage. It reads
    the head's own table first, and otherwise the newest tabled round, which counts only where
    that round names a commit. This takes the table in that same order. `table_shortfall` is
    asked rather than restated, so a partial on record or a cut review history re-reads nothing.
    An empty list counts as unread too. A diff that is really empty costs one schedule of
    re-reads, where an empty list from a failed read decides the coverage.
    """
    if not head_reviews(pr) or head_coverage(pr)[0] != UNSTATED:
        return False
    short = files_shortfall(pr)
    if not short:
        return False
    carried = carried_table(pr)
    named = head_table(pr) or (carried[0] if carried is not None and carried[1] else [])
    return bool(named) and table_shortfall(pr, named) == short


def full_read(owner: str, repo: str, num: int, pr: dict | None = None) -> dict:
    """The full payload, `pr` where one was read already, read again while `files_unread` holds.

    The whole payload is read again rather than the one connection. The digest and the exit code
    then still come from one read of the pull request.
    """
    read = functools.partial(gql, Q_FULL, owner, repo, num)
    return backoff(read() if pr is None else pr, read, files_unread, delays=READ_RETRY_DELAYS)


@functools.cache
def changed_at(owner: str, repo: str, base: str, commit: str) -> frozenset[str] | None:
    """The paths a pull request off `base` changes at `commit`, or None where that was unreadable.

    Read as a three-dot compare, which is the same merge-base reading the pull request's own file
    list is, so the set at an earlier commit and the set at the head are two readings of one
    question rather than two questions. Reading one side from the GraphQL payload instead would
    compare across two APIs with two different ceilings, and a disagreement between them would
    read as the change set having moved.

    None rather than an empty set wherever the read did not happen, since a pull request that
    changes nothing and a compare that could not be fetched take opposite decisions, and the
    caller fails closed on the second. The API caps its file list at 300, and the response carries
    no flag saying it did, so a set at the cap is returned as unreadable rather than as a set that
    is probably short. A name the response leaves null would coerce to the literal path `None`
    and compare equal to another round's null, so an entry that is not a non-empty string makes
    the whole reading unreadable rather than a set with a fabricated member in it.

    A read that did not happen is taken again on the `READ_RETRY_DELAYS` schedule before it is
    returned as None. One run read `unstated` on a head where the runs either side of it carried a
    table. An absence `answered_absent` recognizes is not taken again, being GitHub's own answer.

    Memoized for the run, which is what makes `digest` and `report_verdict` agree. Each computes
    the bound from its own reads, so an unmemoized transient failure in one and not the other
    printed a carried block beside an exit code that refused the carry, and the reverse, which is
    the fail-open direction since the exit code is what a wait keys on. It also reads each ref
    once per run rather than once per caller. It does not close the other window: the two
    compares are still taken at two moments, so a base branch advancing between them still
    manufactures a set mismatch, which refuses the carry and is the fail-closed direction. The
    key is the ref as written, so a branch that advances during a run is answered from the read
    taken before it did, a stale reading held consistently rather than two that disagree.

    The refs are checked before either is interpolated into a compare path, because git permits
    a character that changes the path rather than traveling along it. A `#` truncates it at a
    fragment, answering about a different ref rather than failing. The check refuses that set
    rather than allow-listing the ASCII a branch name usually carries, an allow-list having
    refused a valid `release/1.0+build` and any non-ASCII name. A dot segment is refused
    separately, since it reaches another repository and git forbids one in a ref anyway, so a
    value carrying one did not come from a ref.
    """
    refs = (base, commit)
    if not base or not commit or any(UNSAFE_REF.search(r) or DOT_SEGMENT.search(r) for r in refs):
        return None

    def listed(proc: subprocess.CompletedProcess) -> object:
        if proc.returncode != 0:
            return None
        try:
            return json.loads(proc.stdout.strip())
        except json.JSONDecodeError:
            return None

    compare = functools.partial(
        gh_rest,
        f"repos/{owner}/{repo}/compare/{base}...{commit}",
        r'if (.files | type) == "array" then ([.files[].filename] | @json) else empty end',
    )
    proc = backoff(
        compare(),
        compare,
        lambda p: not answered_absent(p) and listed(p) is None,
        delays=READ_RETRY_DELAYS,
    )
    names = listed(proc)
    if names is None:
        return None
    if not isinstance(names, list) or len(names) >= COMPARE_FILE_CAP:
        return None
    if not all(isinstance(n, str) and n for n in names):
        return None
    return frozenset(names)


def carry_holds(owner: str, repo: str, pr: dict, carried_from: str) -> bool | None:
    """Whether an earlier round's coverage statement still describes this head's change set.

    True only where the set of paths the pull request changes is the same at both commits, which
    is the condition measured on this fleet for the reviewer's own marker carrying: one pull
    request in a promotion chain held its marker across a push whose change set stayed the same
    three files, and the next lost it across a push that grew its change set from one file to
    three. The reviewer keyed on the set rather than on the head moving, and so does this.

    Path identity is the whole of the test, so a push that rewrites the contents of exactly the
    same files carries the statement onto content that statement never read. That is deliberate
    rather than overlooked. The carry is consulted only where a round did cover the head and
    stated nothing, so the reviewer has read this commit and declined to restate its coverage,
    and the delta the digest prints beside the carried reading is what a maintainer judges the
    rest from.

    There is deliberately no short-circuit for a round on the head commit itself. It reads like
    the carry's own case, two rounds on one commit where only the first states coverage, and it
    is unreachable: `head_coverage` takes the worst state over every round on the head and
    `UNSTATED` ranks last in `SEVERITY`, so a head round stating coverage is the head reading and
    neither caller consults the carry at all.

    None where the comparison could not be made at all, which the caller reads as a refusal
    rather than as a pass. A bound that fell open when it could not be measured would be the
    unbounded carry again, reached by a different route.
    """
    head = pr.get("headRefOid") or ""
    base = pr.get("baseRefName") or ""
    earlier = changed_at(owner, repo, base, carried_from)
    current = changed_at(owner, repo, base, head)
    if earlier is None or current is None:
        return None
    return earlier == current


def second_format(body: str) -> bool:
    """Whether this body is written in the second overview format.

    Keyed on the version marker the format opens with, bounded to three spaces of indent, since a
    fourth makes the line a Markdown code block and so a quotation the fence guard never sees.

    Quotations are dropped first for the reason a coverage line's are: this change puts the marker
    into the diff, and a round quoting it is not a round written in that format.
    """
    return bool(CCR_OVERVIEW.search(CODE_SPAN.sub(" ", strip_fences(body or ""))))


def findings_on(tail: str) -> int | None:
    """The total one `**Findings:**` line states, summed over the severities it splits it into.

    None where the line states no total at all, which is the absence of the claim
    `unlisted_findings` subtracts the opened threads from rather than a claim that nothing was
    raised. `None` spelled in words is a
    stated zero, which a round raising nothing writes.

    The markup is masked to one sentinel first, so the digits a badge carries in its own
    attributes are not read as counts. Masking rather than dropping, since dropping joins the
    digits either side into one number. The mask is bounded rather than exhaustive, and the
    four-digit ceiling on a count is the second guard under it: an identifier that survives a
    mask that terminated early is still not read, where an unbounded reading turned one link's
    anchor id into a total of two billion.

    Summed over every count left on the line, because the format states one per severity joined
    by a separator that is presentation rather than structure. Summed rather than paired with the
    markup beside each one, since which side of a count that markup sits on, and whether it is
    markup at all, is the format's presentation rather than its structure: `**3** high` pairs a
    count with no markup and read by pairing alone it summed to nothing.

    Summing overstates rather than understates wherever the line states its counts outside the
    markup, which every body measured does, and which is the direction this picks deliberately.
    A count written inside the markup instead, `[3](url) high`, is masked away with it and the
    line then states no total at all, which prints `?` rather than a fabricated zero.
    A line stating a grand total and then the severities it decomposes into is summed with its
    own parts, and a line stating a total in prose beside a file count sums both. A shortfall
    overstated prints a block a reader checks against the body, where one the reader suppressed
    closes a merge gate over findings nobody answered. `stated_total` takes the largest across
    lines for the same reason.

    `None` is read first and anchored to the front of the line, which is where this format
    states its total. Anchored, so a body naming the word mid-sentence does not fabricate a
    stated zero out of a line stating no total at all, and read first because a line stating the
    zero and then counting something else, `None found in the 3 files reviewed`, states zero
    findings and three files. Read after the counts it reported three findings.
    """
    masked = COUNT_MARKUP.sub(MARKUP_MASK, tail)
    if STATED_NONE.search(masked):
        return 0
    counts = [int(n) for n in COUNT_LOOSE.findall(masked)]
    return sum(counts) if counts else None


DISCUSSION_LINK = re.compile(
    r"\]\((?:https://github\.com/[^/\s()]+/[^/\s()]+/pull/\d+)?#discussion_r(\d+)\)"
)
ENTRY_LINE = re.compile(r"[-*+]\s")
OverviewSection = tuple[str, int | None, list[list[str]], int]


def read_overview(body: str) -> tuple[str, str, list[OverviewSection]]:
    """The verdict, its headline, and the counted sections of one second-format body.

    One reader for every overview section, because each section this format added arrived as a
    spelling nothing here read, and three rounds parked on exit 43 over sections a reader beside
    the others would have answered one at a time. A new spelling is one row in
    `OVERVIEW_SECTIONS`, or one verdict in its set, rather than another reader.

    The verdict is the first verdict heading ahead of the first collapsed section, and the headline
    is the prose under it up to the next heading or bold line, taken from the body unmasked so a
    code span the headline names survives. Each section is a top-level
    `<details>` whose own summary is a row of `OVERVIEW_SECTIONS`, as its role, the count that
    summary states, the thread ids each linked entry links, one list per entry, each id the
    database id of that thread's first comment, and how many of its entries carry such a link.
    An entry is a bullet at the section's margin, outside any block collapsed within it, so an
    indented back-reference or a list an entry collapses is read as neither an entry nor a link.
    One entry can still link a thread beside its own, so a section's links can outnumber its
    linked entries, which is why they are kept per entry. Quotations are masked first, so a body
    quoting a section is not read as carrying one.

    An open section's entries are findings, each checked against the threads, and a resolved
    section's are threads an earlier round raised. The count-first open section also states the
    round's total, having listed that round's own threads in all 17 rounds measured, where
    `Open (N)` listed an earlier round's thread in 2 of 41, so its count is not read as one.
    """
    masked = mask_quotations(body or "")
    tags = mask_quotations(body or "", MARKUP_MASK)
    opener = DETAILS_OPEN.search(tags)
    end = opener.start() if opener else len(masked)
    verdict, headline = "", []
    for line, raw in zip(masked[:end].splitlines(), (body or "")[:end].splitlines()):
        if MARKDOWN_HEADING.match(line):
            if verdict:
                break
            if not unvetted(normal(line), CLEAN_VERDICTS | FLAGGING_VERDICTS):
                verdict = normal(line)
        elif verdict and line.lstrip().startswith("**"):
            break
        elif verdict and line.strip():
            headline.append(raw.strip())
    roles = {k.casefold(): v for k, v in OVERVIEW_SECTIONS.items()}
    sections: list[OverviewSection] = []
    for start, end in details_spans(tags)[0]:
        summary = heading_of(masked[start:end])
        role = roles.get(normal(summary).casefold())
        if role is None:
            continue
        digits = re.search(r"\d+", EMPHASIS.sub("", summary))
        count = counted(digits.group()) if digits else None
        text = masked[start:end]
        own = "\n".join(text[a:b] for a, b in details_spans(tags[start:end])[1])
        entries = [line for line in own.splitlines() if ENTRY_LINE.match(line)]
        ids = [found for line in entries if (found := DISCUSSION_LINK.findall(line))]
        linked = sum(1 for line in entries if DISCUSSION_LINK.search(line))
        sections.append((role, count, ids, linked))
    return verdict, " ".join(headline), sections


def stated_total(body: str, preamble_only: bool = False) -> int | None:
    """The finding total this round states for itself, or None where its overview states none.

    Read from the overview preamble, the text ahead of the first section opener, because a total
    inside a collapsed section is that section's own, the count-first open-findings section below
    being the one exception. Any other total stated only after that opener reads as none stated,
    which prints `?` and no shortfall, so the reading that is missing suppresses the shortfall
    rather than fabricating one.

    The largest wins where the preamble states more than one, so an ambiguous body overstates the
    shortfall rather than suppressing it. The opener is found where each code span is one
    `MARKUP_MASK`, for the reason `details_tags` gives, which keeps every offset the preamble needs.

    A later revision of the format drops the `**Findings:**` line and states its total as a bare
    `**0 open findings**` line, read here the same way, since left unread it printed `?` and
    suppressed the shortfall the total is compared for.

    That revision writes the bare line only for a zero, measured over every body carrying it, and
    states any other total as the count opening its open-findings section, so where the preamble
    states no total, the round's open section states it. `preamble_only` leaves that section out,
    which is how a caller tells which of the two supplied the total.
    """
    stripped = strip_fences(body or "")
    plain = CODE_SPAN.sub(" ", stripped)
    opener = DETAILS_OPEN.search(CODE_SPAN.sub(MARKUP_MASK, stripped))
    preamble = plain[: opener.start()] if opener else plain
    totals = [findings_on(m.group(1)) for m in CCR_FINDINGS.finditer(preamble)]
    totals += [counted(m.group(1)) for m in CCR_OPEN_FINDINGS.finditer(preamble)]
    found = [t for t in totals if t is not None]
    if not found and not preamble_only:
        found = [c for role, c, *_ in read_overview(body)[2] if role == TOTAL and c is not None]
    return max(found, default=None)


def round_threads(pr: dict, review: dict) -> int:
    """How many review threads this round opened, read from the API rather than from its prose.

    The count the stated total is compared against. Six readings tried to take it from the review
    body's own enumeration, and four of them cancelled a shortfall by counting something that was
    not an entry: a link in the change summary, a nested sub-bullet, a back-reference beside an
    entry's own link, a list inside an entry's collapsed detail. Telling those apart needs each
    list marker's own content column and each HTML block's own extent, which is a Markdown parser
    rather than a line scan, and the format supplies no marker that makes the distinction
    structural.

    It is not in the prose to begin with. The anchor each entry links is the database id of the
    thread comment carrying that finding, so the threads the round opened are the same set the
    enumeration names, already structured, and no shape of the body can move the number.

    That equivalence holds only while the enumeration lists what this round raised, and the format
    itself now gives reason to doubt that: a round on this pull request carried a
    `Resolved since last review (N)` section, so the format tracks a finding's state across rounds
    rather than describing one round. If `Open (N)` likewise carries what an earlier round raised
    and this round left open, its total counts findings whose threads belong to that earlier round,
    and this count leaves out exactly those, so it is not compared against the total alone.

    It does: `Open (N)` listed an earlier round's thread in 2 of the 41 rounds `read_overview`
    was measured over, which is why that section's count is not read as a total. The preamble
    total counted those carried findings too, in both rounds measured where a total exceeded the
    threads its round opened. So the comparison counts each open entry standing for an earlier
    round's thread beside `M`, per `carried_open_threads`, rather than this count growing to
    include it, which would move `M` off what the round itself opened.

    A thread beyond the hundred the query reads is not counted, which overstates the shortfall.
    That is the direction that reports, and `threads=` already prints a trailing `+` saying the
    page was cut.
    """
    return len(round_thread_nodes(pr, review))


def round_thread_nodes(pr: dict, review: dict) -> list[dict]:
    """The review threads this round opened, each read by the round its first comment names."""
    rid = review.get("id")
    if not rid:
        return []
    return [
        thread
        for thread in ((pr.get("reviewThreads") or {}).get("nodes") or [])
        for comment in (((thread.get("comments") or {}).get("nodes") or [])[:1])
        if ((comment.get("pullRequestReview") or {}).get("id")) == rid
    ]


def head_overview(pr: dict) -> tuple[int | None, int] | None:
    """The newest round covering the head, as its stated total beside the threads it opened.

    None where no round covers the head and where the one that does is in the first format, so the
    field is absent rather than reading `0/0` on every pull request that has neither.

    Head-scoped for the reason the coverage line is, and unlike the suppressed count: a total is
    one round's statement about one diff, and the push that changes that diff raises a round
    stating its own. The suppressed count is not head-scoped because its blocks carry the finding
    text an answer is owed to, where a shortfall carries a number and nothing to answer.

    The newest round on the head is chosen first and its own body then answers, which is the order
    `review_effort` reads its own field in. Choosing the newest round that happens to be in that
    format instead reports a superseded total as current, and a re-request during the rollout can
    land a round in either format on one commit.
    """
    newest = second_format_head(pr)
    if newest is None:
        return None
    return (stated_total(newest.get("body") or ""), round_threads(pr, newest))


def second_format_head(pr: dict) -> dict | None:
    """The newest round covering the head, where it is written in the second format."""
    newest = newest_of(head_reviews(pr))
    if newest is None or not second_format(newest.get("body") or ""):
        return None
    return newest


def unthreaded_entries(pr: dict) -> int:
    """How many findings the head round's open sections link to no thread on this pull request.

    Each entry links the database id of its thread's first comment, so an id no thread carries is
    a finding named only in the review body, which no thread poll reaches. An entry is read by its
    first link, its title's own anchor, as `carried_open_threads` reads it, so a back-reference
    beside a title anchor that is a thread adds nothing. Distinct title anchors are counted, and a
    resolved section's entries are not, being threads an earlier round raised. A title anchor
    past the hundred threads the query reads counts here too, overstating rather than hiding.
    An `Open (N)` section counting more entries than it has linked entries adds the difference.
    An entry carrying a link this reader does not recognize lands there too, overstating rather
    than hiding.
    """
    known = thread_comment_ids((pr.get("reviewThreads") or {}).get("nodes") or [])
    return len(open_entry_ids(pr) - known) + unlinked_open_entries(pr)


def thread_comment_ids(threads: list[dict]) -> set[str]:
    """The database id of each thread's first comment, the id an overview entry's anchor links."""
    return {
        str(first_comment(thread)["fullDatabaseId"])
        for thread in threads
        if first_comment(thread).get("fullDatabaseId") is not None
    }


def carried_open_threads(pr: dict) -> int:
    """How many of the head round's open entries stand for a thread an earlier round opened.

    A preamble total counts what is still open across rounds rather than what this round raised,
    so a finding an earlier round raised and left open is in the total while its thread is not
    among the ones this round opened. In both rounds measured where a total exceeded the threads
    its round opened, each entry of the open section linked an earlier round's thread and the
    total equalled that section's count. Each such entry
    is counted beside the round's own threads before the two are compared, so a carried-over
    finding with a thread of its own no longer reads as one with none.

    Counted per entry rather than per link, since one entry can link a thread beside its own. An
    entry linking any thread this round opened is that round's finding, already in `M`, so a
    back-reference on it to an earlier thread cancels no shortfall. An entry stands for its
    first link, its title's own anchor, so it is carried only where that link is a thread on this
    pull request. A title anchor naming no thread is `unthreaded_entries`'s to count, and a
    back-reference beside it does not also count the entry here. Two entries standing for one
    earlier thread count once.
    """
    newest = second_format_head(pr)
    if newest is None:
        return 0
    known = thread_comment_ids((pr.get("reviewThreads") or {}).get("nodes") or [])
    own = thread_comment_ids(round_thread_nodes(pr, newest))
    sections = read_overview(newest.get("body") or "")[2]
    return len(
        {
            entry[0]
            for role, _, ids, _ in sections
            if role != RESOLVED
            for entry in ids
            if entry[0] in known and not set(entry) & own
        }
    )


def unlinked_open_entries(pr: dict) -> int:
    """How many entries the head round's `Open (N)` sections count beyond their linked entries."""
    newest = second_format_head(pr)
    if newest is None:
        return 0
    sections = read_overview(newest.get("body") or "")[2]
    return sum(
        max(count - linked, 0)
        for role, count, _, linked in sections
        if role == OPEN and count is not None
    )


def total_section_entries(pr: dict) -> int:
    """How many entries the head round's count-first open-findings sections list with a link.

    Counted per entry rather than per link, since one entry can link a thread beside its own, and
    counting links let that back-reference cancel an entry the section lists with no link at all.
    """
    newest = second_format_head(pr)
    if newest is None:
        return 0
    sections = read_overview(newest.get("body") or "")[2]
    return sum(linked for role, _, _, linked in sections if role == TOTAL)


def open_entry_ids(pr: dict) -> set[str]:
    """The distinct title anchors the head round's open entries link, resolved sections left out."""
    newest = second_format_head(pr)
    if newest is None:
        return set()
    sections = read_overview(newest.get("body") or "")[2]
    return {entry[0] for role, _, ids, _ in sections if role != RESOLVED for entry in ids}


PROSE_FINDINGS = re.compile(
    r"^ {0,3}[*_]{0,2}Review findings[*_]{0,2}:[*_]{0,2}[ \t]*$", re.IGNORECASE
)


def prose_findings(body: str) -> list[list[str]]:
    """The paths each entry of a `Review findings:` list names, one list per entry.

    An entry is a bullet at the margin in the run of lines directly under the label, and an
    indented line continues the entry above it. The label and the bullets are found where
    quotations are masked, so a body quoting the list is not read as carrying one, and each
    entry's paths are then read from its own code spans in the body.
    """
    raw = (body or "").splitlines()
    masked = mask_quotations(body or "").splitlines()
    entries: list[str] = []
    at = 0
    while at < len(masked):
        if not PROSE_FINDINGS.match(masked[at]):
            at += 1
            continue
        at += 1
        while at < len(masked) and not masked[at].strip():
            at += 1
        start = len(entries)
        while at < len(masked):
            line = masked[at]
            if ENTRY_LINE.match(line):
                entries.append(raw[at])
            elif len(entries) > start and line.strip() and line[:1].isspace():
                entries[-1] += "\n" + raw[at]
            else:
                break
            at += 1
    return [
        [
            named
            for span in CODE_SPAN.finditer(entry)
            if (named := bare_path(span.group()[len(span.group(1)) : -len(span.group(1))]).strip())
        ]
        for entry in entries
    ]


def names_path(named: str, path: str) -> bool:
    """Whether a path an entry names is `path`, a trailing part of it, or a directory holding it."""
    named = named.strip("/")
    return bool(named) and (
        path == named or path.endswith("/" + named) or path.startswith(named + "/")
    )


def unthreaded_prose_findings(pr: dict) -> int:
    """How many `Review findings:` entries in the head round name no file its own threads are on.

    Matched by file because an entry names a file and a thread carries one, where its wording
    shares nothing a match could key on. An entry naming no path is counted. An entry whose file
    carries a thread on some other finding is not, which is the case this cannot see.
    """
    newest = second_format_head(pr)
    if newest is None:
        return 0
    paths = {
        bare_path(first_comment(thread).get("path") or "")
        for thread in round_thread_nodes(pr, newest)
    }
    return sum(
        1
        for named in prose_findings(newest.get("body") or "")
        if not any(names_path(n, p) for n in named for p in paths)
    )


def uncounted_verdict(pr: dict) -> tuple[str, str] | None:
    """The head round's verdict and headline, where the verdict flags what nothing counts.

    A round opening on a flagging verdict that states no open finding, lists none in an open
    section, opens no thread, and collapses no suppressed or previously-missed finding names
    whatever it flags in its headline
    alone, which every count here reads as a clean pass. Measured over the rounds in this format,
    the headline named a finding in most such rounds and a caution about the change in the rest,
    and only reading it tells the two apart.

    An open thread defers the reading rather than cancelling it, since the headline may name that
    thread and the loop already waits on it, and once it is resolved the reading returns.
    """
    newest = second_format_head(pr)
    if newest is None:
        return None
    body = newest.get("body") or ""
    verdict, headline, sections = read_overview(body)
    if not verdict or unvetted(verdict, FLAGGING_VERDICTS):
        return None
    listed = any(role != RESOLVED and (count or ids) for role, count, ids, _ in sections)
    if listed or stated_total(body) or round_threads(pr, newest):
        return None
    if suppressed_blocks(body) or previously_missed_blocks(body):
        return None
    if open_threads((pr.get("reviewThreads") or {}).get("nodes") or []):
        return None
    return verdict, headline


def open_threads(threads: list[dict]) -> list[dict]:
    """Every unresolved thread, whoever opened it, since each one blocks a ruleset-gated merge.

    The author decides only how `status` attributes a thread, never whether it counts, because a
    login set read off past history misses the next spelling a reviewer posts under.
    """
    return [t for t in threads if not t.get("isResolved")]


def unlisted_findings(manifest: tuple[int | None, int] | None) -> int:
    """How many findings a round states that it opened no thread for.

    Zero where there is no manifest and zero where it states no total, both being the absence of
    the claim this subtracts from rather than a claim that nothing is missing. Floored at zero,
    since a round that opened more threads than it states findings is the reviewer counting
    differently rather than a negative number of findings.
    """
    if manifest is None or manifest[0] is None:
        return 0
    return max(manifest[0] - manifest[1], 0)


def bare_path(path: str) -> str:
    """The path with its Unicode format characters dropped, the form a table and a diff compare in.

    The second overview format writes a zero-width space after a path's slash, and `str.strip`
    keeps it, so every nested path read as one the diff does not carry while every root-level
    path matched. Both sides are reduced, since a real file name can carry a format character of
    its own, a zero-width joiner in an emoji sequence being one, and reducing only the table's
    side would name that file as omitted.
    """
    return "".join(c for c in path if unicodedata.category(c) != "Cf")


def gap_fits(path: str, parts: list[str]) -> bool:
    """Whether `path` holds `parts` in order, the first opening it, the last closing it.

    At least one character stands between each part and the next. Each middle part is taken at
    its leftmost place, which leaves the most room for the parts after it, so one pass decides it.
    """
    head, *middle, tail = parts
    if not path.startswith(head) or not path.endswith(tail):
        return False
    at, end = len(head), len(path) - len(tail)
    for part in middle:
        found = path.find(part, at + 1, end - 1)
        if found < 0:
            return False
        at = found + len(part)
    return end - at >= 1


def segment_fits(segment: str, pattern: str) -> bool:
    """Whether one path segment matches one row segment, a `[` in either being a literal.

    A segment holding a wildcard must also hold a letter or a digit, so a row of bare wildcards
    such as `*/*` or `*.*` names nothing rather than every path at that depth.
    """
    if not TABLE_GLOB & set(pattern):
        return segment == pattern
    if not any(c.isalnum() for c in pattern):
        return False
    return fnmatch.fnmatchcase(segment, pattern.replace("[", "[[]"))


def row_paths(row: str, diff: set[str]) -> tuple[list[str], bool, bool]:
    """The changed paths a row names, whether it names one at most, and whether a note came off.

    While a row names nothing and ends in a parenthesized note, it is read again without that
    note, since the second format writes one after a path, as in `docs/a.md (cleanup)`. A note
    comes off only while the row names nothing, so a row naming a changed path in any of
    `row_form`'s ways reads as written, whatever parentheses it ends in.

    A row that names its paths only once a note comes off is not counted as naming a path the
    diff does not carry, and it covers none of them either. Nothing reads the note's text, which
    could just as well say the file went unreviewed, so only a row naming a path as written
    covers it.
    """
    noted = False
    while True:
        paths, single = row_form(row, diff)
        if paths or not (head := note_head(row)):
            return paths, single, noted
        row, noted = head, True


def note_split(row: str) -> tuple[str, str]:
    """The row's text before its trailing parenthesized note, and the note, or two empty strings.

    The note runs from the last `(` to the closing `)`, holding no parenthesis of its own, and
    whitespace stands between it and the text before it.
    """
    if not row.endswith(")"):
        return "", ""
    start = row.rfind("(")
    if start < 1 or ")" in row[start + 1 : -1] or not row[start - 1].isspace():
        return "", ""
    return row[:start], row[start:]


def note_head(row: str) -> str:
    """The row before its trailing note, as `note_split` finds one, or empty where it ends in none.

    That text is returned with its outer backticks dropped, as `cell_paths` drops a whole cell's,
    since a code span can come before the note.
    """
    return note_split(row)[0].strip().strip("`").strip()


def row_form(row: str, diff: set[str]) -> tuple[list[str], bool]:
    """The changed paths one table row names, out of `diff`, and whether it names one at most.

    A row names a path in one of three ways, and every claim here and in the digest that a table
    names a set of paths counts all three. A row the diff carries names that path, whatever
    characters it holds. Otherwise a row holding `TABLE_GAP` names every path `gap_fits`, its
    other characters read literally, and since it shortens one path it names one at most.
    Its text outside the gaps must hold a letter or a digit, so a gap between bare punctuation
    names nothing rather than whichever path happens to fit.
    Otherwise a row holding a wildcard names every path it matches one segment at a time, so a
    `*` never crosses a `/`, per `segment_fits`.
    """
    if row in diff:
        return [row], True
    if TABLE_GAP in row:
        parts = row.split(TABLE_GAP)
        if not any(c.isalnum() for c in "".join(parts)):
            return [], True
        return sorted(p for p in diff if gap_fits(p, parts)), True
    if TABLE_GLOB & set(row):
        parts = row.split("/")
        return sorted(
            p
            for p in diff
            if len(segments := p.split("/")) == len(parts)
            and all(map(segment_fits, segments, parts))
        ), False
    return [], True


def table_match(
    named: list[str], diff: set[str]
) -> tuple[list[str], list[str], list[str], list[str]]:
    """The changed paths no row covers, the rows naming none, too many, and the noted paths.

    A row naming one path at most that matches several names none of them, since which one it
    shortened is unknown. A row covers what it names unless a note came off it first, per
    `row_paths`, and the last list holds the uncovered paths such a row names, so it is a subset
    of the first. Each list is sorted, and the table stands in for coverage only where the first
    three are empty.
    """
    covered: set[str] = set()
    noted_paths: set[str] = set()
    invented, ambiguous = [], []
    for row in dict.fromkeys(named):
        paths, single, noted = row_paths(row, diff)
        if not paths:
            invented.append(row)
        elif single and len(paths) > 1:
            ambiguous.append(row)
        else:
            (noted_paths if noted else covered).update(paths)
    left = diff - covered
    return sorted(left), sorted(invented), sorted(ambiguous), sorted(noted_paths & left)


def file_table(body: str) -> list[str]:
    """The paths the round's own file summary table names, in the order it names them.

    Quotations are dropped for the reason a coverage line's are: this change puts a table into
    the diff and a review of it quotes one, and a quoted table read as this round's own names
    files nobody reviewed. That is a fenced block, an unclosed one running to the end of the
    body as Markdown runs it, and a line indented as a code block. Code spans are kept, being how
    the second format writes each path, and a span running over several lines is not masked
    either, since a stray backtick in one row's prose pairs with the next row's path and would
    hide the real table.

    The header is what opens the table and any line that is not a row closes it, so a second
    table later in the body is read as a second table rather than as more of the first.

    Each cell is reduced by `bare_path`, for the reason it states, and read by `cell_paths`.
    """
    paths, reading = [], False
    for line in strip_fences(body or "", to_end=True).splitlines():
        if code_indented(line):
            reading = False
        elif TABLE_HEADER.match(line):
            reading = True
        elif (row := TABLE_ROW.match(line)) is None:
            reading = False
        elif reading and not TABLE_RULE.match(line):
            paths.extend(cell_paths(bare_path(row.group(1))))
    return [p for p in paths if p]


def cell_paths(cell: str) -> list[str]:
    """The paths one table cell names, one per code span where it lists several.

    The second format can group related files into one row, each in a code span of its own with
    a comma between them. Read whole, such a cell is one path holding backticks, which names none
    of the files it lists. So a cell holding single-backtick spans with nothing but commas and
    whitespace outside them names each span.

    Such a list can end in notes, each as `note_split` finds one, and they come off one after
    another as `row_paths` takes them off a row. Each span then keeps those notes, so each reads
    as a noted row, since which file a note is about is unknown. The notes come off before the
    spans are found, so a span quoted inside one is never read as a path. Any other cell is read
    whole, its outer backticks dropped.
    """
    if spans := span_list(cell):
        return [s.strip() for s in spans]
    head, notes = cell.strip(), ""
    while (split := note_split(head))[1]:
        head, notes = split[0].rstrip(), f"{split[1]} {notes}".rstrip()
        if spans := span_list(head):
            return [f"{s.strip()} {notes}" for s in spans]
    return [cell.strip().strip("`").strip()]


def span_list(text: str) -> list[str]:
    """The spans `text` lists, or none where it holds more than spans, commas, and whitespace."""
    spans = TABLE_SPAN.findall(text)
    return spans if spans and TABLE_SEPARATORS.fullmatch(TABLE_SPAN.sub("", text)) else []


def changed_paths(pr: dict) -> tuple[list[str], bool]:
    """The paths this pull request changes, and whether the query's window cut the list short.

    Truncation is carried rather than hidden, since a path outside a short window reads exactly
    like a path the reviewer left out of its table, and the two take opposite readings.
    """
    files = pr.get("files") or {}
    paths = [n.get("path") or "" for n in (files.get("nodes") or [])]
    return [p for p in paths if p], bool((files.get("pageInfo") or {}).get("hasNextPage"))


def head_table(pr: dict) -> list[str]:
    """The file table the rounds covering the current head carry, from the first that has one.

    Any round on the head rather than the one whose counts decided the verdict, since rounds on
    one commit describe one diff, and a re-requested round states the counts again and carries no
    table at all. Thirteen commits here carry more than one round and one of those pairs is that
    shape exactly, so which of the two the verdict happens to read must not decide whether a
    table is found.

    Head-scoped for the reason the coverage line is, and this is the tighter half of the reading:
    three of this repository's four partial pull requests carry their table on the round before a
    push, describing a diff that is not the one being merged, and comparing that table against
    the current changed files would name files as unreviewed on the strength of a stale list.
    Those report as no table rather than as a comparison, which is the honest answer.
    """
    for node in head_reviews(pr):
        named = file_table(node.get("body") or "")
        if named:
            return named
    return []


def table_against_diff(pr: dict, counts: tuple[int, int] | None) -> str:
    """What the reviewer's own file table says about the files the counts leave unread.

    This reports and decides nothing. The exit code stays with the counts, because the table
    tracks them nowhere near well enough to overturn one, and the measurement is what says so.
    Over 348 review bodies on this repository and 121 on ptr727/Blog, the table names the whole
    changed set on partial and fully covered rounds alike: all seven partials on Blog name every
    changed file, as do two of the four here that carry a table. A reading identical under both
    outcomes discriminates neither, which is why a full table is reported as corroborating
    nothing rather than as the count being a miscount.

    The one arm that locates anything is a table shorter by exactly what the counts say went
    unread. One round states 16 of 17 and names 16, omitting `GOVERNANCE.md`, and it is the
    only evidence on record that the unread file is a real file rather than an artifact of
    counting. It stays a lead rather than a verdict, since the table is prose the reviewer
    writes: another states 61 of 62 and names 50, and a third names `GOVENANCE.md`, a path no
    diff here carries.

    A path named that the diff does not carry is what disqualifies the naming arm, that typo
    being enough to drop a real file into the omissions and read it as the one nobody reviewed.
    A shortened path matching several changed files disqualifies it for the same reason, and so
    does a file named only on a row carrying a note, which sits in the omissions although the
    table names it.
    """
    named = head_table(pr)
    if not named:
        return "no round covering this head carries a file table, so none of them locates the file"
    changed, truncated = changed_paths(pr)
    if truncated or not changed:
        return (
            f"the reviewer names {len(named)} entries in its own table and the diff could not "
            f"be read back to compare them, the changed-file list being "
            f"{'longer than the window this reads' if truncated else 'absent from the query'}"
        )
    left, invented, ambiguous, noted = table_match(named, {bare_path(c) for c in changed})
    omitted = [p for p in changed if bare_path(p) in left]
    short = 0 if counts is None else counts[1] - counts[0]
    if not omitted:
        return (
            f"the reviewer's own file table names all {len(changed)} changed files, which it "
            f"also does on rounds stating full coverage, so it corroborates nothing and "
            f"names no unread file"
        )
    if len(omitted) == short and not invented and not ambiguous and not noted:
        return (
            f"the reviewer's own file table omits exactly the {short} file"
            f"{'' if short == 1 else 's'} the counts leave unread, naming "
            f"{', '.join(omitted)}. The table is prose the reviewer writes rather than a "
            f"list from the API, so that is a lead to check rather than a verdict"
        )
    return (
        f"the reviewer's own file table names {len(changed) - len(omitted) + len(noted)} of the "
        f"{len(changed)} changed "
        f"files, omitting {len(omitted) - len(noted)} where the counts leave {short} unread"
        + (f" and naming {', '.join(invented)}, which the diff does not carry" if invented else "")
        + (f" and shortening {', '.join(ambiguous)} to fit several" if ambiguous else "")
        + (
            f" and naming {', '.join(noted)} only on a row with a note, which covers nothing"
            if noted
            else ""
        )
        + ", so it tracks the counts nowhere and names no unread file"
    )


def count_value(digits: str) -> str:
    """A count in a form two counts compare in, leading zeros dropped, and never through `int`.

    `int` refuses a digit run past the interpreter's limit, so a body carrying one would crash
    the digest rather than read as a partial.
    """
    return "".join(str(unicodedata.digit(c)) for c in digits).lstrip("0") or "0"


def body_views(body: str) -> list[str]:
    """The body as each reading of it that `partial_in` scans, every one of them folded.

    Each transform that exposes one hiding place can also hide another: decoding an entity
    can break a tag's quoting, folding a compatibility form can turn a separator into letters,
    and stripping a tag can erase a marker written as one. So no transform replaces another.
    The body is read raw, with entities decoded and format characters dropped, and with
    compatibility forms folded and combining marks dropped too, and each of those with tags
    kept and with tags stripped three ways, a partial found in any one of them counting.
    """
    decoded = "".join(c for c in html.unescape(body) if unicodedata.category(c) != "Cf")
    folded = unicodedata.normalize(
        "NFKC",
        "".join(
            c
            for c in unicodedata.normalize("NFKD", decoded)
            if unicodedata.category(c) not in ("Cf", "Mn")
        ),
    )
    return [
        " ".join(EMPHASIS_MARK.sub("", strip.sub("", base) if strip else base).split())
        for base in (body, decoded, folded)
        for strip in (None, HTML_TAG, LINE_TAG, BARE_TAG)
    ]


def partial_in(view: str) -> str:
    """The first partial marker, marker mention, or count in one reading of a body, or ""."""
    for m in FLEET_REVIEW.finditer(view):
        if count_value(m.group(1)) != count_value(m.group(2)):
            return m.group(0)[:200]
    rest = FLEET_REVIEW.sub(" ", view)
    for name in (MARKER_NAME, MARKER_NAME_TIGHT):
        for m in name.finditer(rest):
            if re.search(r"\d", rest[m.end() : m.end() + MARKER_REACH]):
                return rest[m.start() : m.end() + MARKER_REACH][:200]
    for pattern in (COVERAGE_COUNTS, LOOSE_COUNTS):
        for m in pattern.finditer(view):
            counts = [count_value(g) for g in m.groups() if g is not None][:2]
            if len(counts) == 2 and counts[0] != counts[1]:
                return m.group(0)[:200]
    return ""


def partial_shaped(pr: dict) -> str:
    """The first coverage marker or count in any Copilot round's body that is not a full one.

    Empty where there is none. Read over every round on the pull request, on any commit, with
    no quotation masked, since the table stands in only where no round even appears to have read
    part of a diff. Masking and the line-start anchor are what keep a quoted count from being
    read as a statement, which is right where an unread count blocks as unstated and wrong here,
    where the table would otherwise pass a partial the coverage reader did not recognize.

    Every reading `body_views` gives is scanned, and a partial in any one of them counts, so a
    reading added to catch one shape never hides a shape another reading catches. A well-formed marker counts where its two
    counts differ. Any other mention of the marker's name, in any spelling, counts where a digit
    follows it within `MARKER_REACH` characters, since a marker is written by a model from an
    instruction and a drifted one cannot be parsed reliably, so it is not parsed at all. A mention
    followed by no digit, prose or a placeholder, is skipped. Prose can word a partial in ways no
    pattern here anticipates, so this narrows the gap rather than closing it, which is the cost
    of reading a table at all.
    """
    for node in reviewer_nodes(pr, "reviews"):
        for view in body_views(node.get("body") or ""):
            if found := partial_in(view):
                return found
    return ""


def table_shortfall(pr: dict, named: list[str] | None = None) -> str:
    """Why no round covering the head names exactly this pull request's changed files, or "".

    Empty where the table stands in, which is the coverage reading `table_reading` gives.
    Otherwise the reason, which the digest prints under an unstated head, since the remedies
    differ: a missing or mismatched table may be answered by another round, a changed-file list
    longer than the window this reads by a push bringing it back within it, and a partial on
    record by nothing but the maintainer's reading.

    `named` is an earlier round's table where `table_reading` matches a carried one, and the
    head's own table otherwise. Every other check reads the pull request either way, since a
    partial on record or a cut-short list refuses a carried table exactly as it refuses one on
    the head.
    """
    if reviews_truncated(pr):
        return (
            "the review history is longer than the window this reads, so a round stating "
            "partial coverage may sit out of view"
        )
    if partial := partial_shaped(pr):
        return (
            f"a Copilot round on this pull request states or appears to state partial coverage, "
            f"'{partial}', so the maintainer reads it rather than the table"
        )
    if named is None:
        named = head_table(pr)
    if not named:
        return NO_HEAD_TABLE
    if short := files_shortfall(pr):
        return short
    changed, truncated = changed_paths(pr)
    if truncated:
        return (
            f"the pull request changes more than the {FILES_WINDOW} files this reads, so the "
            f"table cannot be compared against all of them until a push brings the pull request "
            f"back within that window"
        )
    diff = {bare_path(p) for p in changed}
    if len(diff) != len(set(changed)):
        return (
            "two changed paths differ only by a format character, so the table cannot tell "
            "them apart"
        )
    left, invented, ambiguous, noted = table_match(named, diff)
    if not left and not invented and not ambiguous:
        return ""
    omitted = [p for p in left if p not in noted]
    return "the table " + ", and ".join(
        part
        for part in (
            f"leaves out {', '.join(omitted)}" if omitted else "",
            f"names {', '.join(noted)} only on a row with a note, which covers nothing"
            if noted
            else "",
            f"names {', '.join(invented)}, which the diff does not carry" if invented else "",
            f"shortens {', '.join(ambiguous)}, which matches more than one changed file"
            if ambiguous
            else "",
        )
        if part
    )


def carried_table(pr: dict) -> tuple[list[str], str] | None:
    """The newest round on this pull request carrying a file table, its table, and its commit.

    None where no round carries one. A refusal is skipped for the reason `carried_coverage`
    skips it, being a round that read nothing.
    """
    tabled = [
        node
        for node in reviewer_nodes(pr, "reviews")
        if not refusal_of(node) and file_table(node.get("body") or "")
    ]
    newest = newest_of(tabled)
    if newest is None:
        return None
    return file_table(newest.get("body") or ""), (newest.get("commit") or {}).get("oid") or ""


def table_reading(owner: str, repo: str, pr: dict) -> tuple[str, str]:
    """Whether a file table stands in for this head's coverage, as the commit it was read on and why not.

    The shortfall is empty where a table stands in. The commit names the earlier round's commit
    where one was consulted and names one, whether or not its table carries, and is empty
    otherwise.

    This is the coverage reading for a head that no round states coverage on and no earlier
    statement carries to. Copilot's Balanced review effort, the default since 2026-09-28, writes
    the second overview format with a file table and almost never a coverage statement, whatever
    the review instructions ask, so without this reading no pull request reviewed at that effort
    could close its loop.

    It is weaker than a statement, and `table_against_diff` says why: over the rounds measured,
    the table names the whole changed set on partial rounds as well as full ones. So it stands
    in only where nothing on record says any round read part of a diff: no Copilot round on the
    pull request, on any commit, carries a coverage-shaped count that is not a full one, quoted
    or not, and the whole review history is in view. A pull request that ever had a partial
    round goes to the maintainer, since whether that partial still describes this diff is what
    a carry bound refusing, failing, or never reaching it cannot settle. A table naming a file
    the diff does not carry, leaving one out, or shortening a path to fit several is not this
    reading either, and neither is a changed-file list the query cut short or returned malformed.

    A table carries under the bound a statement does, the pull request changing exactly the same
    set of files at both commits, which `carry_holds` reads. Three pull requests in one session
    each had a full table on their first round and a re-review on the next head carrying no
    table, with the file set unchanged between the two, and the maintainer accepted the first
    round's table every time. Only the newest round carrying a table is consulted, and only where
    no round covering the head carries one of its own, since a head table that misses the diff is
    the newer reading and an older one matching does not overrule it.
    """
    shortfall = table_shortfall(pr)
    if shortfall != NO_HEAD_TABLE:
        return "", shortfall
    earlier = carried_table(pr)
    if earlier is None:
        return "", f"{NO_HEAD_TABLE}, and no earlier round carries one either"
    named, carried_from = earlier
    if not carried_from:
        return "", (
            f"{NO_HEAD_TABLE}, and the newest round that does names no commit, so there is no "
            f"change set of its own to compare"
        )
    whose = f"{NO_HEAD_TABLE}, while the newest round that does is on {carried_from[:8]}"
    if why := table_shortfall(pr, named):
        return carried_from, f"{whose}, where {why}"
    kept = carry_holds(owner, repo, pr, carried_from)
    if kept is None:
        return carried_from, (
            f"{whose}, and the change set could not be read at both commits, so its table is "
            f"not carried"
        )
    if not kept:
        return carried_from, (
            f"{whose}, and the pull request changes a different set of files at the two "
            f"commits, so its table describes a diff this head no longer has"
        )
    return carried_from, ""


def badge_text(match: re.Match[str]) -> str:
    """The severity a badge states, taken from whichever quoting its `alt` uses.

    One group per quoting, since a single group cannot hold two alternatives, and the one that
    did not match is `None` rather than empty.
    """
    return match.group(1) if match.group(1) is not None else (match.group(2) or "")


def counted(digits: str) -> int | None:
    """The number this run of digits states, or None where it carries more than a count does."""
    return int(digits) if len(digits) <= COUNT_DIGITS else None


def normal(text: str) -> str:
    """A marker reduced toward what a vetted list compares: emphasis markup dropped, ASCII,
    single spaces, counts as `(N)`.

    Letter case survives, and `unvetted` folds it at the membership test instead.

    The verdict headings carry a colored circle, the suppressed heading carries its finding
    count, and the second overview format wraps each of its own summaries in `<strong>`, so all
    three drift on every review without the section having changed at all.

    A severity badge is reduced to its own `alt` text for the same reason and one more: the markup
    around that text is a pair of versioned icon URLs, so a marker carrying one drifts whenever the
    icon set is rebuilt, and the text it carries is the only part a reader wants.

    A count opening the marker is rewritten the same way, since a later revision of the second
    format writes `3 resolved since last review` where it once closed the summary on `(3)`.
    """
    unbadged = SEVERITY_BADGE.sub(badge_text, text)
    ascii_only = "".join(c for c in EMPHASIS.sub("", unbadged) if ord(c) < 128)
    spaced = re.sub(r"\s+", " ", re.sub(r"\(\d+\)", "(N)", ascii_only)).strip()
    return re.sub(r"^\d+(?= )", "(N)", spaced)


def finding_entry(marker: str) -> bool:
    """Whether this normalized summary is one finding's own heading rather than a section's.

    A section heading is fixed wording a vetted list can hold. A finding entry is a severity and
    then the finding's title, and the title differs on every finding, so no list can ever hold one
    and vetting by text would block on each new finding forever. The severity is what says which of
    the two this is, and `normal` has already reduced the badge carrying it to its own text.

    Recognizing the shape is not the same as passing over the finding. An entry reached this way
    sits inside a collapsed section, and that section is read whole and printed by the digest, so
    what is recognized here is the presentation while the finding itself still reaches a reader.

    A stated count is what keeps that narrow. A section is free to open its own heading with a
    severity, and one recognized on the severity alone would be vetted although nothing here has a
    reader for it, so its findings would reach no digest at all. That is the failure this file
    records three instances of, turned from one that reports itself into one that is silent.
    """
    return bool(SEVERITY_TEXT.match(marker)) and not NORMALIZED_COUNT.search(marker)


def unvetted(marker: str, vetted: set[str]) -> bool:
    """Whether `marker` is absent from `vetted`, compared without regard to letter case."""
    # Folded here rather than in `normal`, whose value the report strings also carry.
    # A report naming a folded shape would not match the body it asks the reader to quote beside it.
    return marker.casefold() not in {v.casefold() for v in vetted}


def unrecognized_in(body: str) -> list[str]:
    """Every marker in one review body this script has no vetted spelling for.

    Read over the body with fenced blocks removed, for the reason the coverage line is: a review
    quoting a heading is not a review carrying one, and this script's own pull requests quote
    these lists in full.

    A body carrying no heading at all is reported rather than passed, since every one of the 333
    measured opens on a heading and a body with none is a format nothing here has seen. A refusal
    is the exemption, being a bare paragraph by design and already classified as one.
    """
    # A refusal is a bare paragraph rather than a review body, and `REFUSAL` is its spelling.
    # It carries no marker to check, and its own opening line is not a coverage statement.
    # Its wording drifting is still caught, since `refusal_of` then stops matching.
    # What is left of a drifted refusal is a body with no heading, which is the arm below.
    if refusal_of({"body": body}):
        return []
    stripped = strip_fences(body or "")
    plain = CODE_SPAN.sub(" ", stripped)
    headings = [normal(ln) for ln in plain.splitlines() if MARKDOWN_HEADING.match(ln)]
    lines = mask_narrative(plain, CODE_SPAN.sub(MARKUP_MASK, stripped)).splitlines()
    labels = [normal(m.group(1)) for m in map(LABEL_LINE.match, lines) if m]
    found = [f"heading: {h}" for h in dict.fromkeys(headings) if unvetted(h, VETTED_HEADINGS)]
    found += [
        f"summary: {marker}"
        # Deduplicated on the raw text before normalizing, which is what it did before the hoist.
        for marker in (normal(s) for s in dict.fromkeys(SUMMARY.findall(plain)))
        if unvetted(marker, VETTED_SUMMARIES) and not finding_entry(marker)
    ]
    found += [
        f"metadata label: {la}" for la in dict.fromkeys(labels) if unvetted(la, VETTED_LABELS)
    ]
    found += [
        f"coverage line: {ln}" for ln in coverage_statements(body) if read_coverage(ln) is None
    ]
    if not headings and not refusal_of({"body": body}):
        found.append("body carrying no heading at all, which no measured review body does")
    return found


def unrecognized_shapes(pr: dict) -> list[str]:
    """Everything about this pull request's reviewer output that this script cannot read.

    Every round rather than the head's, because this asks whether the reader still understands
    the reviewer rather than what the reviewer said about this commit. A shape that arrived one
    round ago is one every later round will carry.
    """
    found = []
    for node in reviewer_nodes(pr, "reviews"):
        where = ((node.get("commit") or {}).get("oid") or "")[:8] or "commit unknown"
        found += [f"{item}  (round {where})" for item in unrecognized_in(node.get("body") or "")]
    return found + reviewer_login_drift(pr)


def reviewer_login_drift(pr: dict) -> list[str]:
    """Logins that read as this reviewer without being the spelling every query here filters on.

    Read from the authors alone, so the liveness query answers it as well as the full one. That
    is what lets the wait stop on a drift rather than poll its whole timeout out against a review
    sitting in plain sight, which is the failure this reading exists to name.
    """
    logins = {
        (n.get("author") or {}).get("login") or ""
        for field in ("reviews", "comments")
        for n in ((pr.get(field) or {}).get("nodes") or [])
    }
    return [
        f"reviewer login: {login}, where every query here filters on {REVIEWER}"
        for login in sorted(logins)
        if login != REVIEWER and READS_AS_REVIEWER.search(login)
    ]


def report_verdict(pr: dict, owner: str, repo: str) -> int:
    """Print the blocking verdict's own status line, and return the exit code it carries.

    The unrecognized shape outranks the coverage one, because a reader that does not understand
    the output cannot be trusted about what it read of the diff either.

    An earlier round's coverage carries forward here as it does in the digest, so the two agree
    on the exit code, and the digest is where the reader is told which round it came from. Which
    round stated coverage is answered by the payload alone, and whether that statement still
    describes this head is not, so this takes the repository and reads the compare `carry_holds`
    bounds the carry on. That reader is memoized for the run, which is what keeps this and the
    digest from reaching two verdicts out of two independent sets of network reads.
    """
    # A coverage line this cannot parse is one of the shapes below rather than a case of its own.
    # It exits here with the remedy that fits it, the reader being what needs the fix.
    if unrecognized_shapes(pr):
        print(
            f"status=UNRECOGNIZED_REVIEWER_OUTPUT this script does not know one or more shapes "
            f"in what the reviewer sent, listed above, so nothing it reports about this review "
            f"is trustworthy and the review loop does not close on this digest. File an issue "
            f"on {HUB} naming each shape and quoting the body it came from, since this script "
            f"is hosted there and the fix lands there. Merging this pull request anyway is the "
            f"maintainer's decision to take and not this script's, and not the agent's."
        )
        return 43
    # A missing head review has no coverage verdict yet.
    # The wait path owns that incomplete state.
    if not head_reviews(pr):
        return 0
    state, line = head_coverage(pr)
    # An earlier round's statement stands until something changes it.
    # The digest printed above this one says which round is carried and what changed since.
    # So this reads the same state and says nothing, two wordings being two places to keep true.
    # `carry_holds` is what bounds it, and it is read here for the same reason the state is.
    # Read as falsy where the bound could not be measured, so an unreadable compare refuses too.
    carried = carried_coverage(pr) if state == UNSTATED else None
    if carried is not None and not carry_holds(owner, repo, pr, carried[2]):
        carried = None
    if carried is not None:
        state, line, _from_head = carried
    if state == UNSTATED and not table_reading(owner, repo, pr)[1]:
        state = TABLE
    if state == PARTIAL:
        # The unread count comes from the line that decided PARTIAL, never from past rounds.
        # None is unreachable there, and narrowing keeps a later change from crashing the gate.
        counts = read_coverage(line)
        if counts is None:
            gap = "that the counts on that line could not be re-read, so the total is unknown"
        else:
            unread = counts[1] - counts[0]
            gap = (
                f"{unread} of the {counts[1]} changed files "
                f"{'has' if unread == 1 else 'have'} no review"
            )
        # The table is read against the counts only where both describe the same diff.
        # A carried partial states counts about an earlier commit, and the table is this head's.
        # The ternary below withholds the reading rather than any refusal `head_table` makes.
        # Withheld although an identical path set means both denominators name the same files.
        # What that earlier round read of those files is still not what this head holds.
        whose = (
            "an earlier round read fewer files than the pull request changed, and the round "
            "covering the head states no coverage of its own, so that reading carries"
            if carried is not None
            else "the review covering the head read fewer files than the pull request changed"
        )
        corroboration = (
            "the counts describe an earlier commit's diff, so the file table on this head is not "
            "read against them"
            if carried is not None
            else table_against_diff(pr, counts)
        )
        print(
            f"status=COVERAGE_IS_PARTIAL {whose}, so part of the diff has no review at all. "
            f"Every partial on "
            f"record stayed partial at the identical ratio across every later round, so a "
            f"re-request is not the remedy it reads as. Splitting is "
            f"the remedy where it applies, and it does not apply to a promotion. Otherwise this "
            f"is the maintainer's call, taken knowing {gap}, and knowing that "
            f"{corroboration}"
        )
        return 42
    if state == UNSTATED:
        print(
            "status=COVERAGE_IS_UNSTATED no statement of changed-file coverage describes this "
            "head, so the review loop cannot prove any round read the full diff. A round that "
            "states none is the ordinary shape of the second overview format and carries the "
            "newest round that states some forward, bounded on the change set, and failing "
            "that, a round covering the head whose own file table names exactly the changed "
            "files stands in, or where none carries a table, the newest round that does, its "
            "table naming exactly the changed files, under the same change-set bound, where no "
            "round on the pull request states or appears to state partial coverage. Reaching here means no table stood in, for the reason the digest "
            "above names, and also one of three things it says which of: no round ever stated "
            "coverage, the round that did describes a different set of changed files than this "
            "head has, or that comparison could not be read. Confirm the head branch carries "
            "the current fleet-code-review skill and Copilot instructions, since a round at Lite "
            "effort states no coverage without them, and where they are missing bring them onto the branch, "
            "push, and run wait, which requests a round on the new head. Where the digest says "
            "the comparison could not be read, run status again, since a failed API read is "
            "one cause of that. Where the digest says the pull request changes more files than "
            "this reads, the table cannot be matched at all, and splitting the pull request is "
            "the remedy where it applies. "
            "A re-request on this same head is not the remedy it reads as, because it returns a "
            "round stating coverage or carrying a table only by chance. Past those, a develop -> "
            "main promotion can still satisfy coverage, in the one case pr-review-conduct's Merge "
            "Gate item 2 states, recording the evidence that item names. Anything else is the "
            "maintainer's call, and merging without coverage is their decision, not the agent's."
        )
        return 45
    if flagged := uncounted_verdict(pr):
        print(
            f"status=VERDICT_NAMES_UNCOUNTED_FINDINGS the round covering the head opens on "
            f"`{flagged[0]}`, a verdict that is not the clean one, while it states no open "
            f"finding, opened no thread, and collapsed no finding, so what it flags is named in "
            f"its headline alone, printed above under VERDICT WITH NO COUNTED FINDING. Give the "
            f"headline the triage a suppressed finding gets: fix what it names and push, or "
            f"answer it on the pull request with `comment`, saying so where it names a caution "
            f"rather than a finding. The answer satisfies this reading, and this code still "
            f"repeats on this head, since nothing here reads an answer, so a caller holding one "
            f"goes on to the other readings"
        )
        return 50
    return 0


def age(stamp: str, now: datetime) -> float | None:
    """Seconds from `stamp` to `now`, or None where the stamp is absent or unparseable.

    None is returned rather than zero or a raise, because every caller treats an unknown age as
    "cannot judge this" and reports nothing. Zero would read as a check that just started, which
    is the reading that reports a job stuck for hours as fresh.
    """
    if not stamp:
        return None
    try:
        return (now - datetime.fromisoformat(stamp)).total_seconds()
    # ValueError is the unparseable stamp, and TypeError is the parseable one carrying no zone.
    # A stamp with no offset parses to a naive datetime, which will not subtract from an aware now.
    # Catching only the first lets that raise out of a reporting call and take the digest with it.
    # A crash is the worst outcome here, since the whole point is to say what the state is.
    except (ValueError, TypeError):
        return None


def head_commit(pr: dict) -> dict:
    """The commit object whose oid is the head's, or an empty dict where none is.

    One reader, because three functions traversed this same shape in three slightly different
    ways and a payload change would have had to be caught in all three. Selection is by oid
    rather than by position, since a rollup off any other commit describes a push ago and renders
    every field, which is a review counted without being read one field along.
    """
    head = pr.get("headRefOid") or ""
    return next(
        (
            c.get("commit") or {}
            for c in ((pr.get("commits") or {}).get("nodes") or [])
            if (c.get("commit") or {}).get("oid") == head
        ),
        {},
    )


def check_nodes(pr: dict) -> list[dict]:
    """The head commit's checks, each a dict with `name`, `state`, `conclusion`, and `since`.

    A readable node also carries `required`, true unless the rollup marks it `isRequired: false`,
    which `held_checks_open` and `wait`'s exit `44` decide on. A node of an unrecognized union
    member carries empty values for the four and an `unreadable` key naming its type instead.

    A rollup carries two node shapes and they spell every field differently: a CheckRun has a
    `name`, a `status` and a `conclusion`, while a StatusContext has a `context` and a single
    `state` that folds both. Reading one shape's keys off the other yields None for all of them,
    which scores an external status as a check in no state at all, so neither pending nor failed.
    They are normalized here so one reading serves both.

    Runs group by workflow and name, or by app and name where the suite has no workflow run.
    Where the workflow is unreadable, an Actions run groups by its own suite and none is dropped.
    Each group keeps every run of its highest suite id and drops its runs from older suites.
    A name only an older suite carries stays, since the merge gate still reads it.
    The highest suite id decides, since `startedAt` misorders overlapping runs.
    A rerun of a superseded suite keeps its older id, so it does not override the newer suite.
    """
    # No match reports nothing rather than falling back to another commit's rollup.
    # A fallback is that same stale reading reached by a different route.
    # The absence is not silent either, and `checks_unreadable` is where the digest says it.
    rollup = head_commit(pr).get("statusCheckRollup") or {}
    out: list[dict] = []
    newest_suite: dict[tuple, int] = {}
    suites: dict[int, tuple[tuple, int]] = {}
    for n in (rollup.get("contexts") or {}).get("nodes") or []:
        if n.get("__typename") == "CheckRun":
            suite = n.get("checkSuite") or {}
            workflow = ((suite.get("workflowRun") or {}).get("workflow") or {}).get("databaseId")
            if workflow:
                origin: tuple = ("workflow", workflow)
            elif (
                pr.get(WORKFLOW_KEYS_UNAVAILABLE)
                and (suite.get("app") or {}).get("slug") == "github-actions"
            ):
                origin = ("suite", suite.get("databaseId"))
            else:
                origin = ("app", (suite.get("app") or {}).get("slug"))
            suite_id = suite.get("databaseId") or 0
            name = n.get("name") or ""
            group = (origin, name)
            newest_suite[group] = max(newest_suite.get(group, 0), suite_id)
            suites[len(out)] = (group, suite_id)
            out.append(
                {
                    "name": n.get("name") or "",
                    "state": n.get("status") or "",
                    "conclusion": n.get("conclusion") or "",
                    "since": n.get("startedAt") or "",
                    "required": n.get("isRequired") is not False,
                }
            )
        elif n.get("__typename") == "StatusContext":
            # A StatusContext reports one field for both, so its state doubles as its conclusion.
            # Its PENDING means the posting system reported the run as under way.
            # A CheckRun's PENDING means the opposite, dispatched and not begun.
            # So the same string is two states, and the shape is knowable only here.
            # Left alone, a long external build reports as queued with no runner assigned.
            # That names a cause the run does not have, on a system that did pick it up.
            state = n.get("state") or ""
            out.append(
                {
                    "name": n.get("context") or "",
                    "state": "IN_PROGRESS" if state == "PENDING" else state,
                    "conclusion": state,
                    "since": n.get("createdAt") or "",
                    "required": n.get("isRequired") is not False,
                }
            )
        else:
            # A third union member is skipped rather than forced into the StatusContext shape.
            # Forcing it reads a label off a node spelling it otherwise, so it renders nameless.
            # It also reads a state that is not there, so it reports as a red check.
            # So skipping is right and skipping *quietly* is not, which review caught here.
            # An unread check missing from the tally is this script's own core failure.
            # The name carries the typename so the digest can say which member it could not read.
            out.append(
                {
                    "name": "",
                    "state": "",
                    "conclusion": "",
                    "since": "",
                    "unreadable": n.get("__typename") or "an unnamed type",
                }
            )
    stale = {i for i, (group, sid) in suites.items() if sid < newest_suite[group]}
    return [node for i, node in enumerate(out) if i not in stale]


def check_shape(node: dict, now: datetime, grace: float, stall: float) -> str:
    """How this check is stuck, or the empty string where it is not.

    Four shapes, because `mergeStateStatus` collapses all of them into `BLOCKED` and each wants
    a different response. A run this session spent twenty-five minutes polling `BLOCKED` on a
    pull request whose only unfinished check was a rollup job no runner ever took, and the cause
    came from the maintainer rather than from any reading here.

    NOT_POSTED is a required status whose poster has not spoken, which is a StatusContext's
    `EXPECTED` and only ever that. It is not a starved job and must not borrow that remedy: no
    runner is owed a status nothing has posted, so re-running a workflow clears nothing, and
    reported as NOT_PICKED_UP it sends a reader at the runner pool over a missing poster.

    NOT_PICKED_UP is the well-founded one: a job GitHub has dispatched but assigned no runner.
    It is read from the queued state rather than from a runner name, which GraphQL does not carry,
    and the state is sufficient because a job held behind a `needs:` dependency does not appear
    in the rollup at all until that dependency finishes, so there is no dependency-blocked queue
    to mistake for a starved one. Observed pickup on a healthy run is one to two minutes, so the
    grace is the pickup grace for the same reason that one is five minutes: inside it, a queued
    job is simply a job starting.

    RUNNING_LONG is deliberately weaker and says so in its wording, because duration alone cannot
    separate a stalled job from a slow one. This repository's own lint job legitimately runs nine
    to eleven minutes while its aggregator is a single shell conditional, and a fleet repository
    building and testing .NET runs longer still. So the threshold is generous, the elapsed time is
    printed for the reader to judge against what the job costs, and nothing here asserts a fault.

    FAILED is any finished check that did not pass, which is a definite answer rather than a stuck
    one, and it is reported here so that a reader is never left deducing a red check from `BLOCKED`.
    """
    state, conclusion = node.get("state") or "", node.get("conclusion") or ""
    elapsed = age(node.get("since") or "", now)
    if state in NOT_POSTED:
        # Its own shape, because the starved remedy is wrong for it in both directions.
        # No runner is owed a status nothing has posted, so re-running a workflow clears nothing.
        # Reported under NOT_PICKED_UP it sent a reader at the runner pool over a missing poster.
        return "NOT_POSTED" if elapsed is not None and elapsed > grace else ""
    if state in NOT_STARTED:
        return "NOT_PICKED_UP" if elapsed is not None and elapsed > grace else ""
    if state == "IN_PROGRESS":
        return "RUNNING_LONG" if elapsed is not None and elapsed > stall else ""
    # A finished check carrying no conclusion yet is still settling, so nothing is reported.
    # Reading an absent verdict as a failure invents a red check out of a race in the API.
    # An unrecognized state would reach the line below and produce that same false failure.
    if not conclusion:
        return ""
    # A conclusion this does not recognize is reported rather than passed over.
    # An unknown verdict blocking a merge is exactly what a reader needs told.
    # A new enum member read as a pass is a red check rendering as a green digest.
    return "" if conclusion in CHECK_OK else "FAILED"


def checks_stuck(
    nodes: list[dict], now: datetime, grace: float, stall: float
) -> list[tuple[dict, str]]:
    """Every check in a stuck shape, as (node, shape), in the order the rollup returns them.

    Takes the normalized list rather than the payload, so one `check_nodes` call serves the tally
    and this together. Two calls parsed the same rollup twice per digest, and this script's whole
    reason for existing is that the reading is the cost.
    """
    return [
        (n, s)
        for n in nodes
        if not n.get("unreadable") and (s := check_shape(n, now, grace, stall))
    ]


def checks_settling(nodes: list[dict], now: datetime, grace: float, stall: float) -> list[dict]:
    """Every check still on its way to a conclusion that waiting can reach.

    A stuck check is left out, since its shape is one no wait clears and polling it only runs the
    timeout out. RUNNING_LONG is the exception, since duration alone cannot tell a stalled job
    from a slow one. Of the rest, a check is settling where it has not passed, which leaves the
    state taxonomy to `check_shape` alone.
    """
    return [
        n
        for n in nodes
        if not n.get("unreadable")
        and check_shape(n, now, grace, stall) in ("", "RUNNING_LONG")
        and (n.get("conclusion") or "") not in CHECK_OK
    ]


def held_checks_open(
    pr: dict, nodes: list[dict], now: datetime, grace: float, stall: float
) -> bool:
    """True where a covered head's checks can still move its merge, so a held wait polls on.

    A merge reading CLEAN, UNSTABLE, or HAS_HOOKS has no required check outstanding, which is
    GitHub's own reading, so on such a merge a check nothing requires does not hold the wait. A
    BLOCKED merge carrying a stuck required check, other than one running long, has its exit, 44,
    decided already, so nothing holds it. Otherwise a required check still settling holds it. So
    does any check still settling while no required check has posted, since a required aggregator
    behind `needs:` enters the rollup only once its dependencies finish. A rollup carrying no check
    at all holds it while the merge reads BLOCKED, BEHIND, or DRAFT, since that is a push whose
    check suites have not registered, where a conflicted, DIRTY one runs no workflow at all. A merge
    reading UNKNOWN holds it whatever the rollup carries, since GitHub has not yet decided the word
    the exit code is chosen by. A pull request no longer open has no merge left to move, and GitHub
    reads UNKNOWN on it for good.
    """
    if pr.get("state", "OPEN") != "OPEN":
        return False
    merge = pr.get("mergeStateStatus")
    if merge in ("CLEAN", "UNSTABLE", "HAS_HOOKS"):
        return False
    if merge == "UNKNOWN":
        return True
    if merge == "BLOCKED" and any(
        n.get("required") and shape != "RUNNING_LONG"
        for n, shape in checks_stuck(nodes, now, grace, stall)
    ):
        return False
    settling = checks_settling(nodes, now, grace, stall)
    if any(n.get("required") for n in settling):
        return True
    if settling and not any(n.get("required") for n in nodes):
        return True
    return not nodes and merge in ("BLOCKED", "BEHIND", "DRAFT") and not checks_unreadable(pr)


def checks_truncated(pr: dict) -> bool:
    """True where the head's rollup carries more contexts than the query asked for.

    The same guard `window_blind` is for, one connection along. A rollup past a hundred contexts
    would drop the rest silently, and a required check among them would be missing from both the
    tally and the stuck reading, so the digest would render a clean pass over a check it never saw.
    That is the exact false clean the rest of this script exists to prevent, and a fleet repository
    with a large matrix build reaches a hundred contexts far sooner than this one does.
    """
    contexts = (head_commit(pr).get("statusCheckRollup") or {}).get("contexts") or {}
    return bool((contexts.get("pageInfo") or {}).get("hasNextPage"))


def checks_unread(nodes: list[dict]) -> list[str]:
    """The rollup typenames this could not read, in the order they appeared.

    A union member that is neither a CheckRun nor a StatusContext cannot be normalized, and it is
    reported rather than dropped. Dropping it is the silent narrowing every other guard here
    exists against: a check absent from the tally and from the stuck reading renders as a clean
    pass over something never seen. Raised in review on this change, against the commit that
    chose skipping over forcing, which was the right half of the answer on its own.
    """
    return [n["unreadable"] for n in nodes if n.get("unreadable")]


def checks_unreadable(pr: dict) -> bool:
    """True where the payload carries commits but none of them is the head.

    `check_nodes` reports nothing in that case rather than reading another commit's rollup, and
    nothing is indistinguishable from a pull request with no checks. So the state is named instead
    of left to render as `checks=0/0`, which a reader would take as a fact about the head rather
    than as this reading having failed. A silent narrowing is the failure mode this whole script
    is built against, and it does not get an exception for its own newest field.
    """
    return bool((pr.get("commits") or {}).get("nodes") or []) and not head_commit(pr)


def checks_tally(nodes: list[dict]) -> tuple[int, int]:
    """(checks that have passed, checks there are), so a bare count says how far the head is."""
    read = [n for n in nodes if not n.get("unreadable")]
    return sum(1 for n in read if (n.get("conclusion") or "") in CHECK_OK), len(read)


def liveness(pr: dict, min_rounds: int = 0) -> tuple[bool, dict | None, list[str]]:
    """Return (head_review_done, answered_outside_review, reviewer_login_drift) off one payload.

    These are the three readings `wait`'s Copilot poll stops on. Its first evaluation and every
    re-read both come here, so a reading added to one cannot be missing from the other. A
    drifted login matches no filter, so `head_review_done` stays false however long the wait
    runs, and waiting it out reports a review that landed as one that never did. `Q_LIVE`
    carries the authors, so the drift costs the poll no extra call. A `min_rounds` of 0 reads
    as `reviewed_head`.
    """
    return head_review_done(pr, min_rounds), answered_outside_review(pr), reviewer_login_drift(pr)


def live_state(
    owner: str, repo: str, num: int, min_rounds: int = 0
) -> tuple[bool, dict | None, list[str]]:
    """Return `liveness` off one fresh `Q_LIVE`, since a push during the wait moves the head."""
    return liveness(gql(Q_LIVE, owner, repo, num), min_rounds)


def heading_of(block: str) -> str:
    """The block's own heading, unwrapped from `<summary>` where it wears one.

    A block starts at its heading, so the heading is the first line whatever markup it carries.
    Reading further would let a finding's own text supply a count the heading never carried, and
    a wrong count reads exactly like a right one.
    """
    head = block.lstrip()
    m = SUMMARY.search(head)
    return m.group(1) if m and head.lower().startswith("<summary") else head.split("\n", 1)[0]


# `<details>(.*?)</details>` lazily pairs each open with the *next* close, which is the innermost one once a shape nests, silently losing everything the outer wrapper still carries after it.
# CodeRabbit's outside-diff section does exactly that: a file wrapper nested inside the section heading, itself wrapping a per-finding "Prompt for AI Agents" block three levels deep.
DETAILS_TAG = re.compile(
    r"(?P<open><details(?:[ \t\n\f\r][^<>]*|/)?>)"
    r"|(?P<lead>(?<![^\n\r]) {0,3}(?P<quote>(?:> {0,4})+)?)?</details"
    r"(?(lead)(?:/|(?:[ \t][^<>\n\r\v\f\x1c\x1d\x1e\x85\u2028\u2029]*)?"
    r"(?(quote)|(?:(?:\r\n|\n|\r)[^<>\n\r\v\f\x1c\x1d\x1e\x85\u2028\u2029]*)?))"
    r"|[ \t]*)>",
    re.IGNORECASE,
)


def details_tags(text: str) -> list[tuple[int, int, bool]]:
    """Each `<details>` tag in `text` as `(start, end, opening)`, starting at its `<` past any lead.

    A close is read anywhere bare. With more before its bracket, or its bracket on the next line, it
    is read only where its own line opens an HTML block: up to three spaces in, behind any
    blockquote markers, its tail allowed one line break where no marker leads it. Behind a marker
    the next line carries its own container, so a tail crossing it would end the close inside
    that line's markup or swallow a line outside the blockquote. Mid-line a next line opening on
    `>` is a blockquote, which leaves the close as text, so a next-line bracket is not read there.
    HTML still reads some closes these rules leave unread, such as one inside an HTML block already
    open. A tail stops at any `<` and at every other line boundary, so it never swallows a later
    tag.

    `text` is read as given, so a caller masks its code spans first, each to `MARKUP_MASK` rather
    than to spaces. A span masked to spaces left the text after it looking indented, and a literal
    mid-line then read as a tag.
    """
    return [
        (max(m.start(), m.end("lead")), m.end(), m.group("open") is not None)
        for m in DETAILS_TAG.finditer(text)
    ]


def details_spans(body: str) -> tuple[list[tuple[int, int]], list[tuple[int, int]]]:
    """Every top-level `<details>...</details>` region's own content as a `(start, end)` span,
    plus the spans of what is left once every top-level region is removed whole.

    Spans rather than text, so one caller can find a heading in one rendering of the body and take
    the block out of another. `marker_blocks` needs exactly that: it looks for a heading in a copy
    with every quotation blanked, and returns the block from the body the review actually carries.
    """
    regions: list[tuple[int, int]] = []
    leftover: list[tuple[int, int]] = []
    depth = 0
    region_start = 0
    cursor = 0
    for start, end, opening in details_tags(body):
        if opening:
            if depth == 0:
                leftover.append((cursor, start))
                cursor = start
                region_start = end
            depth += 1
        elif depth > 0:
            depth -= 1
            if depth == 0:
                regions.append((region_start, start))
                cursor = end
    leftover.append((cursor, len(body)))
    return regions, leftover


def mask_narrative(text: str, tags: str | None = None) -> str:
    """The text with each narrative block's own content blanked, offsets and lines left alone.

    A narrative block is one whose own summary opens it and names `NARRATIVE_SUMMARY`. A block
    nested inside it keeps its content, and an unclosed block is never blanked, so a label it
    holds is still read wherever it sits. The tags are read from `tags` where given, the same text
    with its code spans masked as `details_tags` needs, and everything else from `text`.
    """
    stack: list[tuple[int, int, list[tuple[int, int]]]] = []
    blank: list[tuple[int, int]] = []
    for tag_at, tag_end, opening in details_tags(text if tags is None else tags):
        if opening:
            stack.append((tag_at, tag_end, []))
            continue
        if not stack:
            continue
        tag_start, start, children = stack.pop()
        opener = SUMMARY.match(text[start:tag_at].lstrip())
        if opener and not unvetted(normal(opener.group(1)), {NARRATIVE_SUMMARY}):
            cursor = start
            for child_start, child_end in children:
                blank.append((cursor, child_start))
                cursor = child_end
            blank.append((cursor, tag_at))
        if stack:
            stack[-1][2].append((tag_start, tag_end))
    masked = text
    for start, end in blank:
        masked = masked[:start] + QUOTED_CHAR.sub(" ", masked[start:end]) + masked[end:]
    return masked


# Every character a quotation covers replaced by a space, and every line boundary kept.
# Length-preserving on purpose, so a span found in the masked copy addresses the same text in the body.
# A mask that dropped the quotation instead would shift every offset after it and return a block starting mid-finding.
# The class spells out every character `str.splitlines` breaks on rather than the newline alone, because the block is taken by line index while the span is found by offset.
# A quoted snippet carrying a form feed, a vertical tab, a lone carriage return or a Unicode line separator otherwise left the masked copy one line shorter than the body.
# The block then started a line early and dropped the finding text off its end, which is the defect the spans exist to prevent arriving through the mask instead.
QUOTED_CHAR = re.compile("[^\n\r\v\f\x1c\x1d\x1e\x85\u2028\u2029]")


def mask_quotations(body: str, span: str = " ") -> str:
    """The body with every fenced block and inline code span blanked, offsets left alone.

    A span's characters become `span`, which `details_tags` needs to be other than a space.

    A review quoting a section heading is not a review carrying that section, which is the reading
    every other reader in this file already takes and the one `marker_blocks` did not. This file's
    own test bodies quote these headings, so a round reviewing this script reported the sections
    its test data names as findings the round had raised.
    """
    return CODE_SPAN.sub(
        lambda m: QUOTED_CHAR.sub(span, m.group()),
        strip_fences(body or "", lambda m: QUOTED_CHAR.sub(" ", m.group())),
    )


# The sections that are each other's siblings, so one's heading is where another's block ends.
SECTION_MARKERS = (SUPPRESSED, PREVIOUSLY_MISSED)


def marker_blocks(body: str, marker: re.Pattern[str], strip_blockquote: bool = False) -> list[str]:
    """Return the review body's sections whose own heading matches `marker`, each sliced from
    that heading through to the end of its own region.

    Shared by `suppressed_blocks` and `outside_diff_blocks`. `strip_blockquote` affects heading
    detection only, never the returned content.
    """
    if not body:
        return []
    # Each wrapper's contents, plus what is left outside them all, so a heading is found anywhere.
    # The regions do not overlap, since `details_spans` reports exactly what it removes.
    # Found in the masked copy and taken from the body, which is what makes a quoted heading not a section.
    # The mask keeps every offset, so one set of spans addresses both.
    masked = mask_quotations(body)
    region_spans, leftover_spans = details_spans(mask_quotations(body, MARKUP_MASK))
    blocks = []
    for spans in [[s] for s in region_spans] + [leftover_spans]:
        raw_lines = "".join(body[a:b] for a, b in spans).splitlines()
        # Scanned line by line rather than read from a wrapper's own summary.
        # A heading can sit in the wrapper's body instead, where reading only the summary would miss it.
        scan_lines = "".join(masked[a:b] for a, b in spans).splitlines()
        if strip_blockquote:
            scan_lines = [BLOCKQUOTE.sub("", ln) for ln in scan_lines]
        for i, line in enumerate(scan_lines):
            if marker.search(line) and (HEADING.match(line) or COUNT.search(line)):
                blocks.append("\n".join(raw_lines[i : sibling_heading(scan_lines, i, marker)]))
                break
    return blocks


def sibling_heading(scan_lines: list[str], start: int, marker: re.Pattern[str]) -> int:
    """Where this block ends, which is the next sibling section's heading or the region's end.

    Two sections sharing one wrapper is a shape the corpus carries, a Markdown heading each inside
    the `Review details` wrapper rather than a wrapper of their own. Run to the end of the region
    regardless, the first section's block contains the second whole, and the digest prints those
    findings twice, the second time under a banner asserting the opposite triage about them.

    A heading is not enough to tell the two apart. In this format every finding entry is its own
    `<summary>`, so `HEADING` matches all of them, and the suppressed marker is ordinary English
    that an entry's title carries as readily as a heading does. Bounded on the heading alone, a
    section whose second entry is titled `Low confidence fallback swallows the error` ends at that
    entry, and the entries after it print under the other section's banner, which asserts the
    opposite triage about them, while both counts come out wrong.

    A stated count is what separates them, the same signal `finding_entry` reads: every section
    heading here states one and a finding's own title does not. Requiring it also keeps this off
    the CodeRabbit blocks, whose own marker is in neither of the two and which were never bounded
    before this.

    The cost is named rather than hidden. A sibling heading stating no count does not bound, so a
    block runs into it, which is where this started before the bound existed at all. It is strictly
    narrowing, since the requirement can only lose a bound and never invent one, and every vetted
    heading states a count, so the shape that reaches this also reports `shapes=UNRECOGNIZED` in
    the same digest and tells its reader not to believe the fields beside it.
    """
    for j in range(start + 1, len(scan_lines)):
        line = scan_lines[j]
        if not (HEADING.match(line) and COUNT.search(line)):
            continue
        if any(o is not marker and o.search(line) for o in SECTION_MARKERS):
            return j
    return len(scan_lines)


def suppressed_blocks(body: str) -> list[str]:
    """Copilot's own low-confidence findings, collapsed rather than raised as inline threads.

    See `marker_blocks` for the shared reading.
    """
    return marker_blocks(body, SUPPRESSED)


def unwrap(block: str) -> str:
    """A block reduced to what a reader wants out of it, with its presentation markup dropped.

    Every one of these sections reaches the digest and no thread, so this rendering is the only
    place its findings are read. The wrapper tags say nothing, emphasis around a heading says
    nothing, and a severity badge says one word wrapped in two versioned icon URLs that bury the
    title following them. The word is kept and the markup is not.
    """
    return TAGS.sub("", EMPHASIS.sub("", SEVERITY_BADGE.sub(badge_text, block)))


def previously_missed_blocks(body: str) -> list[str]:
    """Findings the round raises against code that has not changed since the last review.

    The second overview format collapses these into the review body under a counted heading of
    their own, each entry a severity badge and a title, and raises no inline thread for any of
    them. That is a suppressed finding's exact shape, so it is read the same way and for the same
    reason: a finding carrying no thread has no resolved state, and the digest is the only place a
    reader meets it at all.

    Counted across every round rather than the head's, as a suppressed finding is. A round that
    raised one and was then superseded by a push did not thereby have it answered.

    See `marker_blocks` for the shared reading.
    """
    return marker_blocks(body, PREVIOUSLY_MISSED)


def outside_diff_blocks(body: str) -> list[str]:
    """CodeRabbit's own outside-diff-range findings: a real finding on a line outside the pull
    request's changed hunks, which GitHub cannot attach as an inline review comment, so
    CodeRabbit collapses it into the review body instead. Observed corpus: a
    `<summary>...Outside diff range comments (N)</summary>` heading, nested one
    file-level `<details>` deep in turn nesting a per-finding "Prompt for AI Agents" block,
    wrapped in the review's own blockquote.

    Reads a section spanning several files or several findings in one file just as reliably as a
    single finding, since `details_spans` counts the nesting rather than pairing the first
    open it meets with the first close.
    """
    return marker_blocks(body, CR_OUTSIDE_DIFF, strip_blockquote=True)


def finding_count(block: str) -> int:
    """The heading's `(N)`, floored at one, since a block reported as zero reads as a clean pass.

    A run longer than a count carries falls to that same floor rather than crashing the digest,
    the block existing being what says there is at least one finding in it.
    """
    m = COUNT.search(heading_of(block))
    stated = counted(m.group(1)) if m else None
    return max(stated, 1) if stated is not None else 1


def qodo_review_comment(pr: dict) -> dict | None:
    """The reviewer's newest `Code Review by Qodo` comment, where its actual findings live.

    Its formal review object carries an empty body on every review checked, confirmed across
    roughly 60. The findings ride a PR-level comment instead, alongside a second one, `PR
    Summary by Qodo`, that carries none, told apart by this comment's own heading. Comments carry
    no commit, so this is read by recency rather than by head, the same limitation
    `answered_outside_review` already has for Copilot's own plain comments.
    """
    comments = [
        c
        for c in reviewer_nodes(pr, "comments", QODO_LOGIN)
        if QODO_REVIEW_HEADING.search(c.get("body") or "")
    ]
    return max(comments, key=lambda c: c.get("createdAt") or "", default=None)


def qodo_open_findings(body: str) -> list[str]:
    """Each numbered finding in a `Code Review by Qodo` comment that carries neither Qodo's own
    `Resolved` nor `Dismissed` self-tracked badge, so it is still open.

    Qodo nests each finding's own Description/Code/Relevance/Evidence/Agent-prompt sections
    under `<summary>` tags of their own, so every `<summary>` in the comment is read rather than
    only the outermost ones, and only the numbered heading itself, matched by `QODO_FINDING`,
    counts as a finding.
    """
    if not body:
        return []
    return [
        s.strip()
        for s in SUMMARY.findall(CODE_SPAN.sub(" ", strip_fences(body)))
        if QODO_FINDING.match(s) and not QODO_BADGE.search(s)
    ]


def qodo_comments_blind(pr: dict) -> bool:
    """True where Qodo's own `Code Review by Qodo` comment specifically can be sitting behind
    the comments window, so `qodo_open` cannot tell "never commented" from "commented, unseen".

    `window_blind(pr, "comments", QODO_LOGIN)` alone is not enough: it clears the moment any
    Qodo comment is visible, including the paired `PR Summary by Qodo` that carries no findings
    of its own, so a findings comment old enough to fall out of the window would read as though
    Qodo had never posted one at all. A second check closes that gap: an older page exists, some
    Qodo comment is visible, yet `qodo_review_comment` still finds none matching the findings
    heading among what is visible, so the one that matters is the one sitting behind the cut.
    """
    if window_blind(pr, "comments", QODO_LOGIN):
        return True
    older = ((pr.get("comments") or {}).get("pageInfo") or {}).get("hasPreviousPage")
    return (
        bool(older)
        and bool(reviewer_nodes(pr, "comments", QODO_LOGIN))
        and not qodo_review_comment(pr)
    )


def thread_author(t: dict) -> str:
    """The login on a thread's opening comment, or `""` where there is none to read.

    A deleted account leaves `author` present and null, which a chained `.get` would crash on
    without the default here.
    """
    return (((t.get("comments") or {}).get("nodes") or [{}])[0].get("author") or {}).get(
        "login"
    ) or ""


def digest(
    owner: str,
    repo: str,
    num: int,
    seen: set[str] | None = None,
    pr: dict | None = None,
    stalled: str | None = None,
    now: datetime | None = None,
    grace: float = CHECK_GRACE,
    stall: float = CHECK_STALL,
    checks: list[dict] | None = None,
) -> tuple[str, int]:
    """Render the digest from a caller's payload and normalized checks when supplied.

    The caller passes its own readings when the exit code has to agree with what was printed,
    since a review landing between two reads makes a fresh fetch describe a different pull
    request than the one the code was decided from. `checks` spends the rollup parse once,
    since the wait decides an exit code from the reading it just printed.
    `now` is a parameter for the same reason, so a case can hold a check at a known age rather
    than at whatever the clock says when the suite runs.

    `stalled` remains as an ignored compatibility parameter for callers that supplied the old
    pickup reading. GitHub's effort-labeled lifecycle makes that reading inconclusive.
    """
    pr = gql(Q_FULL, owner, repo, num) if pr is None else pr
    now = datetime.now(UTC) if now is None else now
    head = pr["headRefOid"]
    revs = reviewer_nodes(pr, "reviews")
    # `revs` is every round and `on_head` is the ones that reviewed this commit.
    # A refusal sits in the first and not the second, being a round that covered nothing.
    on_head = head_reviews(pr)
    effort, effort_source = review_effort(pr)
    cover, cover_line = head_coverage(pr)
    # An earlier round's statement stands until something changes it.
    # The delta below is what changed, and the maintainer reads the two together.
    # Reassigned rather than held beside `cover`, so the field and the partial block agree.
    # Gated on a round covering the head, since an empty `on_head` reads as unstated too.
    # Ungated, every push carried an earlier round forward onto a head nobody has reviewed.
    # The digest then printed a coverage block beside `review_on_head=NO`.
    # Bounded on the change set, per `carry_holds`.
    # A round on a commit whose diff touched other files clears no gate on a diff nobody has now.
    # The candidate is held apart from the carry, so a refused one is still a round to name.
    # Keyed on the candidate rather than on its commit, which a round can leave unnamed.
    # A review of a commit a force-push made unreachable carries a null `oid`.
    # Keyed on the commit, that round was refused and the digest then said nothing at all.
    candidate = carried_coverage(pr) if cover == UNSTATED and on_head else None
    carried = None
    carried_from, carried_since, carry_kept = "", "", None
    if candidate is not None:
        carried_from = candidate[2]
        carry_kept = carry_holds(owner, repo, pr, carried_from)
        carried_since = delta_since(owner, repo, carried_from, head)
        if carry_kept:
            carried = candidate
            cover, cover_line, _carried_from = candidate
        else:
            carried = None
    table_from, table_short = (
        table_reading(owner, repo, pr) if cover == UNSTATED and on_head else ("", "")
    )
    if cover == UNSTATED and on_head and not table_short:
        cover = TABLE
    local = not on_head and local_cover(pr)
    if local:
        cover = LOCAL
    unknown = unrecognized_shapes(pr)
    threads = pr["reviewThreads"]["nodes"]
    # True where the connection cut off before this pull request's actual thread count.
    # Read here rather than inline below, since both the summary line and the explanatory block need it.
    truncated = threads_truncated(pr)
    # Same reasoning, the `reviews` connection rather than `reviewThreads`, feeding `suppressed=` and `cr_outside_diff=` below.
    revs_truncated = reviews_truncated(pr)
    unresolved = open_threads(threads)
    by_login = {
        login: sum(1 for t in unresolved if thread_author(t) == login) for login in KNOWN_REVIEWERS
    }
    by_login["other"] = sum(1 for t in unresolved if thread_author(t) not in KNOWN_REVIEWERS)
    contributors = {login: n for login, n in by_login.items() if n}
    breakdown = (
        " (" + " ".join(f"{login}={n}" for login, n in contributors.items()) + ")"
        if len(contributors) > 1 or "other" in contributors
        else ""
    )

    # Every round, not just the head, because a suppressed finding has no resolved state to read.
    # Head-scoping treated "superseded by a push" as "answered", and the two are not the same.
    # A finding nobody replied to left the digest the moment the branch moved, reporting zero.
    # That is how four rounds went unanswered across three pull requests in one day.
    # The head is still marked per block, since a finding on an older round may be moot.
    # Deciding that is the reader's call rather than one the count makes for them.
    blocks = [(n, b) for n in revs for b in suppressed_blocks(n.get("body") or "")]
    on_head_blocks = [b for n, b in blocks if (n.get("commit") or {}).get("oid") == head]
    stale = sum(finding_count(b) for n, b in blocks) - sum(finding_count(b) for b in on_head_blocks)
    # Read over the same rounds and split the same way, a previously-missed finding reaching no thread exactly as a suppressed one does.
    # Kept as its own list rather than folded into the one above, because the two take different answers.
    # A suppressed finding is one the reviewer chose to withhold, and this one is a finding it raised outright against code the branch never touched.
    # That one is as likely to be declined as fixed, which is a reader's call rather than one a shared count makes.
    pm_blocks = [(n, b) for n in revs for b in previously_missed_blocks(n.get("body") or "")]
    pm_on_head = [b for n, b in pm_blocks if (n.get("commit") or {}).get("oid") == head]
    pm_stale = sum(finding_count(b) for n, b in pm_blocks) - sum(
        finding_count(b) for b in pm_on_head
    )
    # The second format's stated finding total, beside the threads that round actually opened.
    # It is where that format puts the count the first format's suppressed heading is read for.
    # Absent on a round in the first format, and on a head no round covers.
    # The field below is therefore printed only where there is a round in that format to read.
    manifest = head_overview(pr)
    stated, listed = manifest if manifest else (None, 0)
    unthreaded = unthreaded_entries(pr)
    newest_body = (second_format_head(pr) or {}).get("body") or ""
    sectional = stated is not None and stated_total(newest_body, preamble_only=True) is None
    linked = total_section_entries(pr)
    restated = carried_open_threads(pr)
    prose = unthreaded_prose_findings(pr)
    unlisted = max(
        max((stated or 0) - linked, 0) + unthreaded
        if sectional
        else max(unlisted_findings((stated, listed + restated)), unthreaded),
        prose,
    )
    flagged = uncounted_verdict(pr)

    answer = answered_outside_review(pr)
    refusal = stopping_refusal(pr) or (None if on_head else refusing_review(pr))
    # Read once and handed to the line below, since `quota_refusal` re-walks `refusal_of`.
    refusal_field = (
        "no"
        if not refusal
        else "QUOTA"
        if quota_refusal(refusal) or confirmed_quota(owner, repo, refusal)
        else "ERROR"
        if possible_quota(refusal)
        else "YES"
    )
    blind = [f for f in ("reviews", "comments") if window_blind(pr, f)]
    answered = "yes" if answer else ("unknown" if blind else "no")
    # Normalized once and handed to both readers, since the parse is the cost here.
    # The caller's own list wins where it has one, so the wait parses the rollup once.
    # Without that it read once for what it prints and again for what it returns.
    checks = check_nodes(pr) if checks is None else checks
    ok, total = checks_tally(checks)
    stuck = checks_stuck(checks, now, grace, stall)
    # Which other known reviewers have posted anything at all on this exact head, identity and commit only, no body read.
    # Whether it is a clean pass, a finding, or a refusal of its own is each bot's own prose to parse.
    # Two shapes of that are read below rather than left as each bot's own future task: CodeRabbit's outside-diff-range blocks and Qodo's comment-only findings.
    # Coverage and refusal reading stay Copilot-only, each of the others writing its own findings in its own format.
    # Omitted entirely where none has, so a repository not trialing either stays silent.
    other_on_head = [
        login
        for login in OTHER_REVIEWERS
        if any(
            (n.get("commit") or {}).get("oid") == head for n in reviewer_nodes(pr, "reviews", login)
        )
    ]
    # The service name each tracked other-reviewer's own rate-limit marker gives, where its newest comment or review on this pull request carries one.
    # Not head-scoped, unlike `other_on_head`, since the one observed marker rides a plain comment, which carries no commit to compare against the head at all.
    other_limited = {
        login: name for login in OTHER_REVIEWERS if (name := rate_limited_by(pr, login))
    }
    # CodeRabbit's own outside-diff-range findings, the identical blind spot `blocks` above reads for Copilot.
    # A real finding collapsed into the review body rather than raised as an inline review comment opens no `reviewThreads` entry either.
    # Read every round, not only the head, for the same reason `blocks` is: a finding nobody replied to must not drop out of the digest the moment a later push supersedes its own round.
    cr_revs = reviewer_nodes(pr, "reviews", CODERABBIT_LOGIN)
    cr_blocks = [(n, b) for n in cr_revs for b in outside_diff_blocks(n.get("body") or "")]
    cr_on_head_blocks = [b for n, b in cr_blocks if (n.get("commit") or {}).get("oid") == head]
    cr_stale = sum(finding_count(b) for n, b in cr_blocks) - sum(
        finding_count(b) for b in cr_on_head_blocks
    )
    # Qodo's own comment-only findings: its formal review carries no body at all on any round checked, so its numbered findings are read from its `Code Review by Qodo` PR comment instead.
    # Comments carry no commit, so this is read by recency, not head-scoped the way `cr_blocks` above is.
    qodo_comment = qodo_review_comment(pr)
    qodo_open = qodo_open_findings(qodo_comment.get("body") or "") if qodo_comment else []
    lines = [
        # The repository leads the line, since a number alone reads as correct anywhere.
        # A digest of the wrong pull request is well-formed, so naming it is what shows the miss.
        f"repo={owner}/{repo} pr={num} head={head[:8]} rounds={len(revs)} "
        f"review_on_head={'yes' if on_head else 'local' if local else 'NO'} "
        # Present only where at least one other tracked reviewer has posted on this exact head.
        # No verdict rides on it, unlike `review_on_head`, since nothing here reads what a CodeRabbit or qodo round said, only that one landed.
        + (f"other_reviewed={','.join(other_on_head)} " if other_on_head else "")
        # Present only where a tracked other-reviewer's newest word is that marker.
        # The common case, neither is rate-limited right now, stays silent rather than printing `none`.
        + (
            f"other_rate_limited={','.join(f'{login}:{name}' for login, name in other_limited.items())} "
            if other_limited
            else ""
        )
        + f"effort={effort} effort_source={effort_source} "
        # A field of its own beside that one, since a round can cover the head and read part.
        # Those two readings are what `review_on_head=yes` alone conflates.
        # `carried:` prefixes the state rather than replacing it.
        # What the earlier round said is the reading, and which round said it is the provenance.
        f"coverage={CARRIED + ':' if carried or (cover == TABLE and table_from) else ''}{COVERAGE_FIELD[cover]} "
        # Every other field on this line is a reading of the review.
        # This one says whether the readings can be believed at all, so it is not a count.
        f"shapes={'UNRECOGNIZED' if unknown else 'ok'} "
        # A field of its own, since `rounds=1 review_on_head=NO` is also what a stale round is.
        # The two want opposite responses, one a re-request and the other a split pull request.
        # Upper-case for the reason `NO` is, as a state that blocks a merge is not one to skim.
        # `QUOTA` is its own value rather than folded into `YES`, since its remedy is neither a re-request nor a split.
        # It is the account-level state `status`/`wait` name distinctly.
        f"refusal={refusal_field} "
        # A trailing `+` says the count right before it undercounts, rather than a reader having to notice a fourth field further down to learn the same thing.
        # `unresolved` is drawn from the same truncated `threads` list, so a cut page can hide an open thread exactly as easily as it hides a resolved one, and both counts carry the marker.
        f"threads={len(threads)}{'+' if truncated else ''} "
        f"unresolved={len(unresolved)}{'+' if truncated else ''}{breakdown} "
        # A trailing `+` on either count says the same as it does on `threads=`/`unresolved=` above: a round old enough to fall out of the `reviews` window is a round its own finding cannot be read from.
        f"suppressed={sum(finding_count(b) for n, b in blocks)}{'+' if revs_truncated else ''} "
        f"(on_head={sum(finding_count(b) for b in on_head_blocks)} earlier={stale}) "
        # Present where the round raised at least one finding against unchanged code, or the truncated window means one could exist unseen.
        # Silent otherwise, rather than printing a permanent zero on every repository whose reviewer never raises one.
        # No exit code rides on it, the same as `suppressed=`: it reports a finding to triage rather than a state that blocks.
        + (
            f"previously_missed={sum(finding_count(b) for n, b in pm_blocks)}{'+' if revs_truncated else ''} "
            f"(on_head={sum(finding_count(b) for b in pm_on_head)} earlier={pm_stale}) "
            if pm_blocks or revs_truncated
            else ""
        )
        # Present only where the round covering the head is written in the second format.
        # `?` where that round states no total, which is not the same reading as a total of zero.
        + (f"overview={'?' if stated is None else stated}/{listed} " if manifest else "")
        # Present where CodeRabbit has raised at least one outside-diff-range finding on any round, or the truncated window means one could exist unseen.
        # A repository not trialing it, on an untruncated window, stays silent rather than printing a permanent `cr_outside_diff=0`.
        + (
            f"cr_outside_diff={sum(finding_count(b) for n, b in cr_blocks)}{'+' if revs_truncated else ''} "
            f"(on_head={sum(finding_count(b) for b in cr_on_head_blocks)} earlier={cr_stale}) "
            if cr_blocks or revs_truncated
            else ""
        )
        # Present once Qodo has posted a `Code Review by Qodo` comment at all, `0` included.
        # A `0` here is itself a reading, Qodo reviewed and left nothing open, rather than silence about whether it reviewed.
        # `unknown` where its findings comment specifically can be sitting behind the window, so its silence here cannot be told apart from it never having commented.
        # Absent only where `qodo_comments_blind` clears it and it genuinely never has.
        + (
            f"qodo_open={len(qodo_open)} "
            if qodo_comment
            else ("qodo_open=unknown " if qodo_comments_blind(pr) else "")
        )
        + f"answered_outside_review={answered} "
        f"requested={'yes' if reviewer_requested(pr) else 'no'} "
        f"merge={pr.get('mergeStateStatus')} "
        # `merge=BLOCKED` names no cause and is worn by every gate alike.
        # A red check, a check nothing runs, an open thread, and a missing approval all read it.
        # One run here polled that word for twenty-five minutes without learning which it was.
        # The tally says how far the head got, and `checks=0/0` says the rollup carried nothing.
        f"checks={ok}/{total}"
        # Named only where there is one to name.
        # A field reading `none` on every green run is one a reader skips when it finally says more.
         + (f" stuck={','.join(sorted({s for _, s in stuck}))}" if stuck else "")
    ]
    if refusal:
        # Printed whole for the reason the comment below is, as the wording carries the remedy.
        # A file-count refusal is cleared by splitting the pull request.
        # A quota refusal is account-level state nothing here clears, so the caller proceeds on the other reviewers' coverage.
        # This reads neither cause, only that the round declined.
        lines.append(
            "  COPILOT REFUSED THIS ROUND: the newest Copilot review on this pull request says "
            "it did not review, so it covers nothing and re-requesting repeats it, and the body "
            "below is what says which remedy applies"
        )
        lines += [
            f"    {ln.rstrip()}" for ln in (refusal.get("body") or "").splitlines() if ln.strip()
        ]
        if limit := confirmed_quota(owner, repo, refusal):
            lines.append(f"  RATE LIMIT FROM THE REVIEWER'S RUN LOG: {limit}")
    if unknown:
        # First of the blocks, since it says how far the rest of them can be trusted.
        lines.append(
            f"  UNRECOGNIZED REVIEWER OUTPUT ({len(unknown)}): the shapes below are ones "
            "this script has no reader for, so every other field here is a reading of "
            "output it does not fully understand and a clean digest does not mean a "
            f"clean review. File an issue on {HUB}, which hosts this script, naming each "
            "shape and quoting the body it came from, before closing the review loop. "
            "Whether to merge anyway is the maintainer's call rather than the agent's"
        )
        lines += [f"    {item}" for item in unknown]
    if unlisted:
        # The count is all this shape carries, unlike a suppressed block, whose text prints.
        # What the finding says is in the review body, which is where this sends the reader.
        lines.append(
            f"  FINDINGS WITH NO THREAD ({unlisted}): the round covering the head states "
            + ("no total" if stated is None else f"{stated} finding{'' if stated == 1 else 's'}")
            + (
                f", its open section listing {linked} with a link"
                if sectional
                else f" and opened {listed} thread{'' if listed == 1 else 's'}"
                + (
                    f", its open section also linking {restated} "
                    f"thread{'' if restated == 1 else 's'} an earlier round opened"
                    if restated
                    else ""
                )
            )
            + (
                f", and {unthreaded} of the entries its open sections count "
                f"{'has' if unthreaded == 1 else 'have'} no title link this script reads to a "
                "thread on this pull request"
                if unthreaded
                else ""
            )
            + (
                f", and {prose} of the entries under its `Review findings:` label "
                f"{'names' if prose == 1 else 'name'} no file a thread of that round is on"
                if prose
                else ""
            )
            + f", so {unlisted} "
            f"{'finding is' if unlisted == 1 else 'findings are'} raised where polling threads "
            "cannot see them. Read the review body itself and give each the triage a suppressed "
            "finding gets, before closing the review loop. A thread past the hundred this reads "
            "would count here too, and `threads=` carries a trailing `+` where that page was cut"
        )
    if flagged:
        lines.append(
            f"  VERDICT WITH NO COUNTED FINDING: the round covering the head opens on "
            f"`{flagged[0]}` and counts no finding, so its headline is the one place what it "
            "flags is named. Give the headline the triage a suppressed finding gets"
        )
        lines.append(f"    {flagged[1] or '(the headline carries no prose)'}")
    if carried is not None:
        # Printed whatever the carried state is, the provenance being the same question either way.
        # A carried partial then prints its own block under this one.
        lines.append(
            f"  COVERAGE IS CARRIED: the round covering the head states none, which is the "
            f"ordinary shape of the second overview format rather than a defect in it. The "
            f"newest round that states any is {carried_from[:8] or 'a commit this cannot name'}, "
            f"and {carried_since}. Whether that round still covers this head is the "
            f"maintainer's reading, and the delta is what to read it from"
        )
        lines.append(f"    {cover_line}")
    elif candidate is not None:
        # The reader is told which round it declined and on what ground.
        # The alternative is a digest printing nothing about a statement it found and did not use.
        if not carried_from:
            why = "that round names no commit, so there is no change set of its own to compare"
        elif carry_kept is None:
            why = (
                "the change set could not be read at both commits, so the bound could not be "
                "measured and the statement is not carried"
            )
        else:
            why = (
                "the pull request changes a different set of files at the two commits, so "
                "that statement describes a diff this head no longer has"
            )
        lines.append(
            f"  COVERAGE IS NOT CARRIED: the newest round that states any is "
            f"{carried_from[:8] or 'a commit this cannot name'}, and {carried_since}, but {why}"
        )
    if cover == TABLE:
        count = len(changed_paths(pr)[0])
        files = f"{count} changed file{'' if count == 1 else 's'}"
        lines.append(
            "  COVERAGE IS READ FROM THE FILE TABLE: no round states coverage of this head and "
            "none carries to it, and "
            + (
                f"no round covering it carries a file table, so the newest round that does, on "
                f"{table_from[:8]}, carries its table, which names exactly the {files}, the "
                f"pull request changing that same set at both commits, and "
                f"{delta_since(owner, repo, table_from, head)}"
                if table_from
                else f"a round covering it names exactly the {files} in its own table"
            )
        )
    elif cover == UNSTATED and on_head:
        lines.append(f"  NO FILE TABLE STANDS IN: {table_short}")
    elif cover == LOCAL:
        lines.append(
            "  COVERAGE IS READ FROM THE LOCAL PASS: no Copilot round covers this head, which "
            "follows the pull request's first round, and a comment from an owner, member, or "
            "collaborator attests a recorded local strict-review pass over exactly this head's "
            f"content, which recorded {attestation(pr)} finding(s)"
        )
    if cover == PARTIAL:
        # The line prints under the marker for the reason a suppressed block does.
        # The counts say how much of the diff went unread, and no thread carries them.
        lines.append(
            "  COVERAGE IS PARTIAL: "
            + (
                "an earlier round read fewer files than the pull request changed and the round "
                "covering the head states none, so that reading carries"
                if carried is not None
                else "the review covering the head read fewer files than the pull request changed"
            )
            + ", so files in the diff have no review at all. A "
            "re-request has never cleared this on record, so report it rather than "
            "retrying into it, and let the maintainer take the merge decision"
        )
        lines.append(f"    {cover_line}")
        # The counts decide the state and this line says what the reviewer's file table adds.
        # Printed under the same marker, since a maintainer taking the decision wants both.
        # It names no verdict of its own, having been measured against the counts it sits under.
        # The table tracks them nowhere, which is why it is reported rather than read.
        # Withheld on a carried partial, the counts and the table then describing two diffs.
        lines.append(
            "    the counts describe an earlier commit's diff, so the file table on this head "
            "is not read against them"
            if carried is not None
            else f"    {table_against_diff(pr, read_coverage(cover_line))}"
        )
    for node, shape in stuck:
        # Each shape carries its own remedy, which is the whole point of telling them apart.
        # A reader handed one word for all three retries the wrong thing, or waits on a queue.
        elapsed = age(node.get("since") or "", now)
        # Clamped for display only, since a machine clock behind GitHub's renders `queued -3m`.
        # A negative age reads as nonsense and hides how long the thing has actually waited.
        # The comparison above keeps the raw value, so skew cannot make a stuck check report.
        mins = "age unknown" if elapsed is None else f"{int(max(elapsed, 0) // 60)}m"
        # Read with `.get` for the reason `age` catches two exceptions.
        # A caller handing this an odd node shape should cost a field, never the whole digest.
        # Reporting the state is the one job here, so it must not be the thing that raises.
        name = node.get("name") or "unnamed"
        if shape == "NOT_PICKED_UP":
            lines.append(
                f"  CHECK NOT PICKED UP ({name!r}, queued {mins} with no "
                "runner assigned): nothing here starts it, because the runner pool is "
                "GitHub-hosted, so re-run the workflow or wait on that capacity"
            )
        elif shape == "NOT_POSTED":
            lines.append(
                f"  CHECK NEVER POSTED ({name!r}, expected {mins} and not "
                "reported): a required status whose poster has not spoken, so no runner "
                "is owed it and re-running a workflow here clears nothing"
            )
        elif shape == "RUNNING_LONG":
            lines.append(
                f"  CHECK RUNNING LONG ({name!r}, running {mins}): it has a "
                "runner, so it is not starved, and whether this is hung or merely slow "
                "is a judgment against what this job normally costs"
            )
        else:
            lines.append(
                f"  CHECK FAILED ({name!r}, {node.get('state')}/"
                f"{node.get('conclusion')}): a verdict rather than a stuck check, so "
                "read the run and fix what failed, since no wait and no re-run clears "
                "a real failure"
            )
    unread = checks_unread(checks)
    if unread:
        lines.append(
            f"  CHECKS PARTIALLY UNREAD ({', '.join(sorted(set(unread)))}): the rollup "
            "carries a context type this cannot normalize, so it is absent from the "
            "tally and the stuck reading alike and neither speaks for it"
        )
    if checks_truncated(pr):
        lines.append(
            f"  CHECKS TRUNCATED: the head carries more than the {CHECKS_WINDOW} contexts "
            "this query asks for, so a check past the window is missing from the tally "
            "and from the stuck reading alike, and neither reports what it did not see"
        )
    if checks_unreadable(pr):
        lines.append(
            "  CHECKS UNREADABLE: the payload carries commits and none of them is the "
            "head, so no rollup here describes this head and `checks=0/0` is this "
            "reading failing rather than a pull request with no checks"
        )
    if pr.get(WORKFLOW_KEYS_UNAVAILABLE):
        lines.append(
            "  CHECKS DEDUPED WITHOUT WORKFLOW KEYS: `workflowRun` was unreadable, "
            "so each Actions run is kept per suite and a superseded failure may still "
            "show, which reads red rather than hiding a failing check"
        )
    if blind:
        lines.append(
            f"  BEHIND THE WINDOW ({' and '.join(blind)}): the newest {WINDOW} carry "
            "none from the reviewer and older ones exist, so this cannot decide"
        )
    if truncated:
        lines.append(
            "  THREADS TRUNCATED: this pull request carries more review threads than the "
            "100 read here, and the connection reads oldest-first, so the ones cut off are "
            "the newest rather than the oldest, exactly the ones most likely to still be "
            "open. `threads=` and `unresolved=` above undercount. Read the rest with "
            "`reply`'s own paginated walk before trusting either number"
        )
    if revs_truncated:
        lines.append(
            "  REVIEWS TRUNCATED: this pull request carries more reviews than the 100 read "
            "here, and the connection reads newest-first, so the ones cut off are the "
            "oldest, exactly where an unanswered suppressed or outside-diff-range finding is "
            "most likely to still sit unresolved. `suppressed=` and `cr_outside_diff=` above "
            "undercount"
        )
    if answer:
        # Printed whole for the same reason a suppressed finding is, since it reaches no thread.
        # Its wording is the only thing separating a refusal from an ordinary remark.
        lines.append(
            f"  COPILOT COMMENT ({answer.get('createdAt')}, newer than any review): "
            "the reviewer answered without reviewing, so read the body below and "
            "decide, since a refusal is terminal and a remark is not"
        )
        lines += [
            f"    {ln.rstrip()}" for ln in (answer.get("body") or "").splitlines() if ln.strip()
        ]
    new = 0
    for t in unresolved:
        c = (t.get("comments") or {}).get("nodes", [{}])[0]
        tid = t["id"]
        mark = ""
        if seen is not None:
            if tid in seen:
                continue
            seen.add(tid)
            mark = "NEW "
            new += 1
        body = " ".join((c.get("body") or "").split())
        lines.append(f"  {mark}{tid} {c.get('path')}:{c.get('line')} {body[:160]}")
    for n, b in blocks:
        # Printed whole where a thread body is truncated, since a thread can be re-read at its id.
        # A suppressed finding has none, so this digest is the only place it appears.
        # GraphQL returns a null commit for a pending or partial review.
        # An empty sha rendered as "raised on , earlier round", losing what traces the finding.
        sha = ((n.get("commit") or {}).get("oid") or "")[:8]
        if not sha:
            where = "commit unknown, treat as outstanding"
        elif sha == head[:8]:
            where = "on head"
        else:
            where = f"raised on {sha}, earlier round"
        lines.append(
            f"  SUPPRESSED ({where}): no thread to resolve, "
            "answer it in the PR conversation quoting the finding"
        )
        # Indentation is kept, since a block carries fenced code a flattened line would garble.
        lines += [f"    {ln.rstrip()}" for ln in unwrap(b).splitlines() if ln.strip()]
    for n, b in pm_blocks:
        # Mirrors the `SUPPRESSED` rendering above, its own reasons applying identically here.
        sha = ((n.get("commit") or {}).get("oid") or "")[:8]
        if not sha:
            where = "commit unknown, treat as outstanding"
        elif sha == head[:8]:
            where = "on head"
        else:
            where = f"raised on {sha}, earlier round"
        lines.append(
            f"  PREVIOUSLY MISSED ({where}): no thread to resolve, "
            "raised against code this branch did not change, "
            "answer it in the PR conversation quoting the finding"
        )
        lines += [f"    {ln.rstrip()}" for ln in unwrap(b).splitlines() if ln.strip()]
    for n, b in cr_blocks:
        # Mirrors the `SUPPRESSED` rendering above, its own reasons applying identically here.
        sha = ((n.get("commit") or {}).get("oid") or "")[:8]
        if not sha:
            where = "commit unknown, treat as outstanding"
        elif sha == head[:8]:
            where = "on head"
        else:
            where = f"raised on {sha}, earlier round"
        lines.append(
            f"  CODERABBIT OUTSIDE-DIFF ({where}): no thread to resolve, "
            "answer it in the PR conversation quoting the finding"
        )
        lines += [f"    {ln.rstrip()}" for ln in unwrap(b).splitlines() if ln.strip()]
    for finding in qodo_open:
        # Not head-scoped, since the comment it comes from carries no commit at all.
        lines.append(
            "  QODO OPEN FINDING: neither Resolved nor Dismissed, no thread to resolve, "
            "answer it in the PR conversation quoting the finding, and spot-verify Qodo's "
            "own badge against `gh pr diff` rather than trusting it outright"
        )
        lines.append(f"    {' '.join(finding.split())}")
    if seen is not None:
        lines[0] += f" new={new}"
    return "\n".join(lines), len(unresolved)


def origin_owner() -> str | None:
    """The owner `git remote get-url origin` yields against this script's own directory, or None.

    Anchored on the script's own directory rather than the working directory, because this is
    reached from a hub checkout while the repository being answered is named on the command line,
    so the working directory says nothing about who owns either.
    """
    # An inherited GIT_DIR or GIT_COMMON_DIR overrides the `-C` argument and points the probe at another repository's config.
    # An inherited GIT_CONFIG* channel injects config into this call, and `remote get-url` applies `insteadOf` rewriting, so an injected rewrite would make it name an owner this checkout does not have.
    # This strip closes the channels an inherited tooling environment sets: the four discovery names it drops, and any name starting with GIT_CONFIG that injects config into this one call.
    # It does not close, and cannot close, an inherited HOME, XDG_CONFIG_HOME, or PATH, because a caller who controls any of those already controls the process that would perform the write, leaving nothing here for this check to defend.
    # The environment is stripped rather than cleared, since git still needs PATH to be found at all, and HOME so the checkout's own global config and `safe.directory` still apply.
    env = {
        k: v
        for k, v in os.environ.items()
        if k not in ("GIT_DIR", "GIT_WORK_TREE", "GIT_COMMON_DIR", "GIT_OBJECT_DIRECTORY")
        and not k.startswith("GIT_CONFIG")
    }
    try:
        url = subprocess.run(
            ["git", "-C", str(HERE), "remote", "get-url", "origin"],
            capture_output=True,
            text=True,
            encoding="utf-8",
            timeout=5,
            check=False,
            env=env,
        ).stdout.strip()
    # A missing git, a timeout, or any other failure all mean the owner cannot be read.
    except Exception:  # noqa: BLE001
        return None
    m = re.search(r"[:/]([A-Za-z0-9_.\-]+)/([A-Za-z0-9_.\-]+?)(?:\.git)?/?$", url)
    return m.group(1).lower() if m else None


def in_scope(target_owner: str) -> tuple[bool, str]:
    """Whether writing to `target_owner` is in scope here, with the reason where it is not.

    Same owner covers the origin and every sibling, which is the whole fleet. A different owner is
    the shape that once put a comment on a stranger's pull request, and it is refused outright
    rather than granted by an environment variable, because a variable this process can be handed
    is one the caller can set on the command that runs it, and a grant the caller writes for
    itself is not a grant. The `gh-write-guard` hook reads that grant from the session it was
    launched with, which is why the cross-owner case belongs on the runbook's explicit `gh` path
    where the hook is the one adjudicating it.
    """
    origin = origin_owner()
    if origin is None:
        return False, (
            "this script got no owner from `git remote get-url origin` against the directory "
            f"it sits in, {HERE}, so the owner a write would stay within cannot be "
            "established, and an unverified scope is not a scope. That directory is this "
            "script's own rather than the working directory, so a readable `origin` in the "
            "working directory says nothing about this, and the remedy for a copy of this "
            "script sitting outside a checkout is to run the copy that lives inside a hub "
            "checkout instead of this one"
        )
    if target_owner.lower() != origin:
        return False, (
            f"the target is under {target_owner}, and `git remote get-url origin`, run against "
            f"the directory this script sits in, {HERE}, named {origin}. A different owner is "
            "the shape this refuses outright: take it through the runbook mutations, where the "
            "write-guard hook reads the maintainer grant"
        )
    return True, ""


def first_comment(thread: dict) -> dict:
    """The thread's opening comment, which is the finding itself."""
    return ((thread.get("comments") or {}).get("nodes") or [{}])[0]


def unresolved_threads(owner: str, repo: str, num: int) -> list[dict]:
    """Every unresolved review thread, following the cursor to the end.

    Stopping at the first page reports no match on a thread that is merely further along, and a
    no-match is indistinguishable from a thread that was already answered.
    """
    out: list[dict] = []
    after = None
    while True:
        extra: dict[str, str] = {"after": after} if after else {}
        conn = gh_graphql(Q_THREADS, o=owner, r=repo, n=num, **extra)["repository"]["pullRequest"][
            "reviewThreads"
        ]
        out += open_threads(conn["nodes"])
        page = conn.get("pageInfo") or {}
        if not page.get("hasNextPage"):
            return out
        after = page["endCursor"]


def describe(thread: dict) -> str:
    """One line naming a thread by what a reader recognizes it as, never by its id.

    The 120-character cutoff below is a display limit rather than a matching one: `--match`
    itself compares against the whole body, so a pattern copied from past the cutoff has nothing
    printed to copy from, and widening the words or adding `--path` is the way past that, not a
    longer cutoff here.
    """
    c = first_comment(thread)
    body = " ".join((c.get("body") or "").split())
    return (
        f"{thread.get('path')}:{thread.get('line')} "
        f"by {(c.get('author') or {}).get('login')}: {body[:120]}"
    )


_TYPOGRAPHIC_FOLD = str.maketrans(
    {
        "\u2018": "'",  # left single quotation mark
        "\u2019": "'",  # right single quotation mark
        "\u201c": '"',  # left double quotation mark
        "\u201d": '"',  # right double quotation mark
        "\u2013": "-",  # en dash
        "\u2014": "-",  # em dash
        "\u2026": "...",  # horizontal ellipsis
    }
)


def matching_threads(threads: list[dict], match: str, path: str | None) -> list[dict]:
    """Threads whose finding text contains `match`, narrowed by `path` where one is given.

    Matched on the finding's own words rather than on a line number, because a fix push moves the
    line and every lookup keyed to one then misses: replies posted against nothing while the
    resolves still succeeded, so the threads closed carrying no answer. Case-insensitive, since
    the text is quoted back out of a digest by a reader rather than compared by a machine. Both
    sides are folded through `_TYPOGRAPHIC_FOLD` first, since the pattern is a substring copied
    from a rendered finding whose typographic quotes, dashes, or ellipsis may not survive that
    copy in ASCII. Both sides also fold whitespace to a single space, the same collapse
    `describe()` applies before printing a thread's body as an `unresolved:` line, so a pattern
    copied from that printed line still matches the raw body it was copied from, whether the
    whitespace the printed line flattened was a line break, a tab, or a run of spaces.
    """
    needle = fold_finding(match)
    return [
        t for t in threads if needle in finding_text(t) and (path is None or t.get("path") == path)
    ]


def fold_finding(text: str) -> str:
    """The form `--match` compares in, as `matching_threads` describes."""
    return " ".join(text.split()).translate(_TYPOGRAPHIC_FOLD).lower()


def finding_text(thread: dict) -> str:
    """A thread's opening comment in the form `--match` compares against."""
    return fold_finding(first_comment(thread).get("body") or "")


def comment_on_pr(owner: str, repo: str, num: int, body: str) -> int:
    """Post one PR conversation comment and confirm the returned comment. Returns an exit code."""
    ok, why = in_scope(owner)
    if not ok:
        print(f"status=OUT_OF_SCOPE nothing was written: {why}")
        return 64

    target = gql(Q_COMMENT_TARGET, owner, repo, num)
    if not target or not target.get("id") or not target.get("url"):
        print(
            f"status=TARGET_NOT_READ nothing was written: {owner}/{repo} #{num} did not "
            "return a pull request id and URL, so the mutation has no verified target"
        )
        return 65

    body = body.replace("\r\n", "\n").replace("\r", "\n")
    edge = (gh_graphql(M_COMMENT, subjectId=target["id"], body=body).get("addComment") or {}).get(
        "commentEdge"
    ) or {}
    comment = edge.get("node") or {}
    if not comment.get("url") or (comment.get("body") or "") != body:
        print(
            "status=COMMENT_NOT_CONFIRMED the response returned no URL or a different body. "
            "Inspect the PR conversation before retrying, since the write may have landed."
        )
        print(f"  response: {json.dumps(comment)[:400]}")
        return 66
    print(f"commented: {comment['url']}")
    print("status=COMMENTED")
    return 0


def attest(owner: str, repo: str, num: int, checkout: str) -> int:
    """Publish that a recorded local pass covers this pull request's head. Returns an exit code.

    The receipt `local_review.py` records lives in the checkout's git directory, so nothing on
    GitHub can read it, and the review gate for a fix push needs to. This reads the receipt where it
    lives and posts a comment the gate can read, and only after four checks, each of which would
    otherwise vouch for content the pull request does not carry: the checkout's HEAD is the pull
    request's head, the checkout holds no change beyond that commit, its merge base with the base
    branch is the pull request's own, and `local_review.py status` reports a current pass against
    the pull request's base, one read answering the coverage and the findings count alike. The first
    two are read again once the check returns, since the checkout can move while it runs.
    """
    ok, why = in_scope(owner)
    if not ok:
        print(f"status=OUT_OF_SCOPE nothing was written: {why}")
        return 64
    target = gql(Q_ATTEST_TARGET, owner, repo, num) or {}
    head, base = target.get("headRefOid") or "", target.get("baseRefName") or ""
    if not head or not base:
        print(
            f"status=TARGET_NOT_READ nothing was written: {owner}/{repo} #{num} did not return "
            "its head commit and base branch"
        )
        return 65
    local = _git(checkout, "rev-parse", "HEAD")
    dirty = _git(
        checkout, "status", "--porcelain", "--untracked-files=all", "--ignore-submodules=none"
    )
    if local.returncode != 0 or dirty.returncode != 0:
        print(
            f"status=CHECKOUT_NOT_READ nothing was written: {checkout} is not a readable checkout"
        )
        return 67
    if local.stdout.strip() != head:
        print(
            f"status=CHECKOUT_NOT_THE_HEAD nothing was written: the checkout is at "
            f"{local.stdout.strip()[:8]} and the pull request's head is {head[:8]}, so push or "
            "fetch until they agree"
        )
        return 67
    if dirty.stdout.strip():
        print(
            "status=CHECKOUT_NOT_THE_HEAD nothing was written: the checkout holds changes the "
            "head commit does not, so a pass over it does not describe what was pushed"
        )
        return 67
    if UNSAFE_REF.search(base) or DOT_SEGMENT.search(base):
        print(
            f"status=MERGE_BASE_NOT_READ nothing was written: the base branch {base!r} carries a "
            "character that would change the compare path rather than travel along it"
        )
        return 67
    ours = _git(checkout, "merge-base", f"refs/remotes/origin/{base}", "HEAD")
    theirs = gh_rest(
        f"repos/{owner}/{repo}/compare/{base}...{head}", ".merge_base_commit.sha // empty"
    )
    if ours.returncode != 0 or theirs.returncode != 0 or not theirs.stdout.strip():
        print(
            f"status=MERGE_BASE_NOT_READ nothing was written: the merge base with {base} could "
            "not be read in the checkout or from GitHub, so the pass's scope cannot be compared "
            "with the pull request's"
        )
        print(f"  {(ours.stderr or theirs.stderr).strip()[:400]}")
        return 67
    if ours.stdout.strip() != theirs.stdout.strip():
        print(
            f"status=CHECKOUT_NOT_THE_HEAD nothing was written: the checkout's merge base with "
            f"origin/{base} is not the pull request's own, so a pass against it measured a "
            "different change set. Fetch, or attest from a checkout of the pull request's own "
            "repository"
        )
        return 67
    covered, findings, said = receipt_reading(checkout, base)
    if not covered:
        print(
            f"status=NO_LOCAL_PASS nothing was written: `local_review.py status --target {base}` "
            "reports no recorded pass covering this content. Run the local strict review, "
            "record it, and attest again"
        )
        print(f"  {said.strip()[:400]}")
        return 68
    after = _git(checkout, "rev-parse", "HEAD")
    still = _git(
        checkout, "status", "--porcelain", "--untracked-files=all", "--ignore-submodules=none"
    )
    if after.stdout.strip() != head or still.returncode != 0 or still.stdout.strip():
        print(
            "status=CHECKOUT_NOT_THE_HEAD nothing was written: the checkout moved or changed "
            "while the pass was being checked, so the check no longer describes the head"
        )
        return 67
    body = (
        f"A recorded local strict-review pass covers head `{head}`, the content this pull "
        f"request carries at that commit against `{base}`, and it recorded {findings} "
        f"finding{'' if findings == '1' else 's'}.\n\n"
        f"<!-- fleet-local-review: head={head} base={base} findings={findings} -->"
    )
    return comment_on_pr(owner, repo, num, body)


def receipt_reading(checkout: str, base: str) -> tuple[bool, str, str]:
    """Whether a recorded pass covers the checkout's content, its findings, and what was read.

    One `local_review.py status` read answers both, so the coverage and the count describe the
    same receipt, where a check followed by a second read could straddle a record replacing it.
    The findings are the covering passes' counts summed, or "unknown" where any recorded none,
    and they are shown rather than gated, since a local pass's findings are advisory.
    """
    try:
        proc = subprocess.run(
            [sys.executable, str(LOCAL_REVIEW), "status", "--target", base],
            cwd=checkout,
            capture_output=True,
            text=True,
            encoding="utf-8",
            timeout=120,
            check=False,
        )
    except (OSError, subprocess.SubprocessError):
        return False, "unknown", "local_review.py could not be run"
    try:
        data = json.loads(proc.stdout)
    except ValueError:
        return False, "unknown", proc.stdout or proc.stderr
    if proc.returncode != 0 or not isinstance(data, dict) or data.get("covered") is not True:
        return False, "unknown", proc.stdout or proc.stderr
    if data.get("receiptProblems"):
        return False, "unknown", proc.stdout
    counts = data.get("findings")
    if (
        not isinstance(counts, dict)
        or not counts
        or not all(isinstance(n, int) and n >= 0 for n in counts.values())
    ):
        return True, "unknown", proc.stdout
    return True, str(sum(counts.values())), proc.stdout


def _git(checkout: str, *args: str) -> subprocess.CompletedProcess:
    """One read-only git command in `checkout`, returned whole rather than raised."""
    try:
        return subprocess.run(
            ["git", "-C", checkout, *args],
            capture_output=True,
            text=True,
            encoding="utf-8",
            timeout=30,
            check=False,
        )
    except (OSError, subprocess.SubprocessError):
        return subprocess.CompletedProcess([], 1, "", "git could not be run")


def attested(pr: dict) -> bool:
    """Whether an attestation vouches for this pull request's head, per `attestation`."""
    return attestation(pr) is not None


def attestation(pr: dict) -> str | None:
    """The findings count an owner's, member's, or collaborator's attestation of this head carries.

    None where no attestation vouches for the head against the pull request's current base, so
    a retargeted pull request needs a new one, and "unknown" where the pass recorded no count.

    Read over every comment rather than the reviewer's own, since an attestation is the
    maintainer's account speaking. The association is what keeps a passer-by's comment carrying
    the same marker from vouching for anything. The marker counts only as a line of its own
    outside a fence, so a comment quoting it in a span or a code block vouches for nothing.
    """
    head, base = pr.get("headRefOid") or "", pr.get("baseRefName") or ""
    for node in (pr.get("comments") or {}).get("nodes") or []:
        if (node.get("authorAssociation") or "") not in TRUSTED_ASSOCIATIONS:
            continue
        body = CODE_SPAN.sub(" ", strip_fences(node.get("body") or "", to_end=True))
        for m in ATTESTATION.finditer(body):
            if m.group(1) == head and m.group(2) == base:
                return m.group(3)
    return None


def promotion(pr: dict) -> bool:
    """Whether the pull request merges into the repository's default branch.

    That is the promotion, whose own Copilot round is the backstop reading the whole diff, so
    the local pass never stands in for it. A default branch the payload does not name reads as
    a promotion too, since the failure that way is one more Copilot request rather than a gate
    passed on a local pass alone.
    """
    default = ((pr.get("baseRepository") or {}).get("defaultBranchRef") or {}).get("name") or ""
    return not default or pr.get("baseRefName") == default


def first_round_done(pr: dict) -> bool:
    """Whether Copilot has reviewed this pull request at all, on any head, a refusal not counting."""
    return any(not refusal_of(n) for n in reviewer_nodes(pr, "reviews"))


def holds(pr: dict) -> bool:
    """Whether this head is a fix push the local pass covers, so `wait` requests no round for it.

    Read again on the payload the verdict is graded on, since a push between two reads moves the
    head the hold was decided for.
    """
    return (
        not promotion(pr)
        and first_round_done(pr)
        and not reviews_truncated(pr)
        and not partial_shaped(pr)
        and not reviewer_requested(pr)
    )


def local_cover(pr: dict) -> bool:
    """Whether a recorded local pass stands in for a Copilot round on this head.

    Only on a pull request into a branch other than the default, after Copilot's first round, on
    a head its writer attests, and only where nothing on record says any round read part of a
    diff and the whole review history is in view, the bound a file table stands in under. A fix
    push after the first round is covered by the local pass, and a promotion keeps its own
    Copilot round. A refusal on this head is Copilot's own word on it and is never overruled.
    """
    return (
        not refusing_review(pr)
        and not promotion(pr)
        and first_round_done(pr)
        and attested(pr)
        and not reviews_truncated(pr)
        and not partial_shaped(pr)
    )


POLL_DELAYS = (15, 20, 30, 45, 60, 120)


def backoff[Polled](
    value: Polled,
    read: Callable[[], Polled],
    keep: Callable[[Polled], bool],
    start: float | None = None,
    timeout: float = 0.0,
    delays: Iterable[float] | None = None,
) -> Polled:
    """Re-read `value` on a schedule of delays while `keep` holds and the bound allows.

    `delays` defaults to `POLL_DELAYS`, its last delay repeated only where a clock bounds the loop,
    so no call runs unbounded. A finite schedule ends the loop by running out, with no clock where
    `start` is None. Otherwise `start` is a `time.monotonic` reading and `timeout` is in seconds.
    The bound is checked before each sleep, so the poll can run past it by up to one delay.
    Returns the last value read, or `value` itself where the poll never sleeps.

    One loop for every re-read here, since a fix applied to one inline copy and not another
    makes one wait time out differently from another.
    The backoff runs in-process, so the whole wait costs one agent turn.
    """
    if delays is None:
        repeated = itertools.repeat(POLL_DELAYS[-1]) if start is not None else ()
        delays = itertools.chain(POLL_DELAYS, repeated)
    for delay in delays:
        if not keep(value) or (start is not None and time.monotonic() - start > timeout):
            break
        time.sleep(delay)
        value = read()
    return value


def reply_to_thread(
    owner: str,
    repo: str,
    num: int,
    match: str,
    body: str,
    path: str | None,
    resolve: bool,
    all_identical: bool = False,
) -> int:
    """Answer the one thread `match` selects, and resolve it where asked. Returns an exit code.

    Every refusal below is a stop rather than a fallback. There is no id to guess at, no
    second-best thread to settle for, and no resolve on a reply that did not land, because each
    of those closes a finding while leaving it unanswered, which is the state a reviewer reads as
    addressed. A no-match names the unresolved count, since zero and several otherwise read the
    same without the reader counting the lines the refusal prints below it.

    `all_identical` covers the one ambiguity no wording can narrow: a reviewer posting the same
    finding as two threads. Their opening comments fold to the same text, so every pattern that
    selects one selects the other, and the refusal left both open with no route to either. The
    flag answers every candidate with the one body, and only when they all sit on one file and
    fold identically, so two findings that merely share the quoted words still refuse, and the
    same text on two files stays two findings that `--path` already tells apart.
    """
    ok, why = in_scope(owner)
    if not ok:
        print(f"status=OUT_OF_SCOPE nothing was written: {why}")
        return 64

    threads = unresolved_threads(owner, repo, num)
    hits = matching_threads(threads, match, path)
    if not hits:
        print(
            f"status=NO_MATCH nothing was written: no unresolved thread on {owner}/{repo} "
            f"#{num} carries {match!r}"
            + (f" at {path}" if path else "")
            + f", of {len(threads)} unresolved thread(s) total"
            + ". Widen the words or drop --path rather than reaching for an id, since the "
            "thread may also be resolved already, which reads the same from here."
        )
        for t in threads:
            print(f"  unresolved: {describe(t)}")
        return 60
    paths = {t.get("path") for t in hits}
    identical = len(paths) == 1 and len({finding_text(t) for t in hits}) == 1
    if len(hits) > 1 and not (all_identical and identical):
        if len(paths) > 1:
            hint = "The candidates sit on different files, so add --path to select one."
        elif identical:
            hint = (
                "Every candidate opens with the same text, so no wording selects one: pass "
                "--all-identical to answer each of them with this body."
            )
        elif all_identical:
            hint = (
                "--all-identical answers only candidates whose opening comments read the same, "
                "and these differ, so quote more of the finding."
            )
        else:
            hint = "Quote more of the finding, or add --path."
        print(
            f"status=AMBIGUOUS nothing was written: {len(hits)} unresolved threads carry "
            f"{match!r}, and picking one of them is the failure this avoids rather than a "
            f"default it can take. {hint}"
        )
        for t in hits:
            print(f"  candidate: {describe(t)}")
        return 61

    for done, target in enumerate(hits):
        code = answer_thread(target, body, resolve)
        if code:
            if len(hits) > 1:
                replied = done + (code == 63)
                print(
                    f"  {replied} of {len(hits)} identical threads are confirmed to carry this reply, "
                    "and none after this one was attempted."
                )
            return code

    count = f" ({len(hits)} identical threads)" if len(hits) > 1 else ""
    if not resolve:
        print(
            f"status=REPLIED{count} answered and left open, since --resolve was not given. "
            "A decline is resolved only once its evidence is in the thread."
        )
        return 0
    print(f"status=REPLIED_AND_RESOLVED{count}")
    return 0


def answer_thread(target: dict, body: str, resolve: bool) -> int:
    """Reply to one selected thread and resolve it where asked. Returns 0, or an exit code."""
    print(f"answering: {describe(target)}")
    reply = (
        gh_graphql(M_REPLY, threadId=target["id"], body=body).get("addPullRequestReviewThreadReply")
        or {}
    )
    comment = reply.get("comment") or {}
    # The url is what says a reply carried a body, and an empty one still returns a comment.
    # Resolving past that closes the thread with nothing in it, which is what happened three times.
    if not comment.get("url") or not (comment.get("body") or "").strip():
        print(
            "status=REPLY_NOT_CONFIRMED the reply returned no url or an empty body, so the "
            "thread is NOT resolved and the answer is not recorded. Read the response above "
            "before retrying, since a write that appears to fail may have taken on the server."
        )
        print(f"  response: {json.dumps(reply)[:400]}")
        return 62
    print(f"replied: {comment['url']}")

    if not resolve:
        return 0

    thread = (gh_graphql(M_RESOLVE, threadId=target["id"]).get("resolveReviewThread") or {}).get(
        "thread"
    ) or {}
    if not thread.get("isResolved"):
        print(
            "status=RESOLVE_NOT_CONFIRMED the reply landed and the resolve did not report the "
            "thread resolved, so it is still open and the answer is already posted."
        )
        print(f"  response: {json.dumps(thread)[:400]}")
        return 63
    return 0


def gh_rest(path: str, jq: str | None = None, raw: bool = False) -> subprocess.CompletedProcess:
    """One REST read, returned whole so the caller can tell an absent object from an unread one.

    Unlike `gh_graphql` this does not raise on a non-zero exit, because a 404 here is an answer
    the caller acts on rather than a failure. Reads only: every path passed in is a GET.

    `raw` is for a job log, which carries terminal escape sequences that `gh` 2.97 and later
    refuse to print without `--allow-escape-sequences`. An older `gh` has no such flag and
    prints the log as it is, so a run refused on the flag is retried without it.
    """
    base = ["gh", "api", path] + (["--jq", jq] if jq else [])
    proc = _gh_run(base + (["--allow-escape-sequences"] if raw else []), raw)
    if raw and proc.returncode != 0 and "unknown flag: --allow-escape-sequences" in proc.stderr:
        proc = _gh_run(base, raw)
    return proc


def _gh_run(argv: list[str], raw: bool) -> subprocess.CompletedProcess:
    """Run one `gh` read, a log decoded leniently since its bytes are the runner's own."""
    try:
        return subprocess.run(
            argv,
            capture_output=True,
            text=True,
            encoding="utf-8",
            errors="replace" if raw else "strict",
            timeout=30,
            check=False,
        )
    except (OSError, subprocess.SubprocessError):
        return subprocess.CompletedProcess(argv, 1, "", "gh could not be run")


def answered_absent(proc: subprocess.CompletedProcess) -> bool:
    """Whether GitHub said the object is not there, as opposed to nothing having been read."""
    m = HTTP_STATUS.search(proc.stderr)
    return m is not None and m.group(1) in ABSENT


def body_references(body: str) -> tuple[list[str], list[str]]:
    """The `uses:` refs and the claimed commits a description quotes, de-duplicated and sorted.

    A commit counts only where a verb claims this branch carries it. A SHA mentioned without one
    is a mention of history, of another repository, or of quoted output, and the corpus above says
    that is what a description's SHAs almost always are.

    Prose claims stay out for the same reason the free SHA scan did. Judging those needs a
    similarity heuristic, which `spec/section-model.md` rejects.

    A claimed SHA still has to carry a digit, as a backstop on the vocabulary above: `accede` and
    `defaced` inflect into all-hex English words, so a verb this list gains later cannot start
    reading one of them as a commit. It costs about one real SHA in a thousand.
    """
    uses = sorted({m.group("ref") for m in BODY_USES.finditer(body)})
    shas = sorted(
        {
            m.group("sha")
            for m in BODY_CLAIM.finditer(body)
            if any(c.isdigit() for c in m.group("sha"))
        }
    )
    return uses, shas


def commit_state(owner: str, repo: str, sha: str, head: str) -> str:
    """The empty string where `head` carries `sha`, `UNREAD` where nothing was read, else why not.

    Ancestry rather than membership of the branch's own commits, so a description may cite a
    commit it inherited from the base branch. GitHub reports the comparison from the base's side,
    so `identical` and `ahead` are the two readings that say the head carries it.
    """
    proc = gh_rest(f"repos/{owner}/{repo}/compare/{sha}...{head}", ".status")
    if proc.returncode == 0:
        status = proc.stdout.strip()
        if status in ("identical", "ahead"):
            return ""
        return (
            f"{owner}/{repo} carries the commit and this head does not descend from it ({status})"
        )
    if answered_absent(proc):
        return f"{owner}/{repo} carries no such commit"
    return UNREAD


def head_carries(owner: str, repo: str, head: str, refs: list[str]) -> set[str] | None:
    """Which of `refs` appear anywhere in the tree at `head`, or None where nothing was read.

    Read at the head commit rather than from a local checkout, because the branch being described
    need not be fetched here and a working tree is not what a description is measured against.

    The whole tree rather than a guessed set of workflow paths: this repo carries `uses:` lines in
    catalog snippets and in documentation as well as under `.github/`, and a surface narrower than
    the tree would report a ref as absent because it looked in the wrong place. One archive is
    also one request, where walking a listing costs a request per file and grows with the repo.
    """
    # Read as bytes, so this cannot go through `gh_rest` and carries that helper's guards itself.
    # An absent `gh` or a hung download reads as undecided, like every other unreadable answer.
    # Raising instead would abort a run the caller is in the middle of.
    try:
        proc = subprocess.run(
            ["gh", "api", f"repos/{owner}/{repo}/tarball/{head}"],
            capture_output=True,
            timeout=TARBALL_TIMEOUT,
            check=False,
        )
    except (OSError, subprocess.SubprocessError):
        return None
    if proc.returncode != 0:
        return None
    needles = {r: r.encode() for r in refs}
    found: set[str] = set()
    try:
        with tarfile.open(fileobj=io.BytesIO(proc.stdout), mode="r:gz") as archive:
            for member in archive:
                if not member.isfile():
                    continue
                handle = archive.extractfile(member)
                if handle is None:
                    continue
                # Matched as bytes, so a file this cannot decode is searched rather than skipped.
                # No encoding guess can then turn a present ref into an absent one.
                blob = handle.read()
                found |= {r for r, n in needles.items() if r not in found and n in blob}
                if len(found) == len(needles):
                    break
    except (tarfile.TarError, OSError, EOFError):
        return None
    return found


def check_claims(owner: str, repo: str, num: int) -> int:
    """Report every reference the description makes that its own head tree does not carry."""
    pr = gql(Q_CLAIMS, owner, repo, num)
    head, body = pr["headRefOid"], pr.get("body") or ""
    uses, shas = body_references(body)

    stale, unread = [], 0
    for sha in shas:
        state = commit_state(owner, repo, sha, head)
        if state == UNREAD:
            unread += 1
        elif state:
            stale.append(
                f"STALE COMMIT `{sha}`: {state}, so the description points at something "
                "this branch does not have"
            )

    # One read for every ref together, since the archive it reads is the same one either way.
    carried = head_carries(owner, repo, head, uses) if uses else set()
    for ref in uses:
        if carried is None:
            unread += 1
        elif ref not in carried:
            stale.append(f"STALE USES `{ref}`: no file at this head carries that ref")

    print(
        f"repo={owner}/{repo} pr={num} head={head[:8]} commits={len(shas)} uses={len(uses)} "
        f"stale={len(stale)} unread={unread}"
    )
    for line in stale:
        print(f"  {line}")
    if stale:
        # Named as the description's problem rather than the branch's.
        # The branch is the ground truth here, and the body is what drifted away from it.
        print(
            "status=DESCRIPTION_CONTRADICTS_ITS_BRANCH update the body to what the head tree "
            "carries, since a reference that resolves to nothing is caught by a reviewer or "
            "not at all"
        )
        return 70
    if unread and unread == len(shas) + len(uses):
        # A check that decided nothing prints the same `stale=0` a clean one does.
        print(
            "status=NOTHING_WAS_READ every reference in the description was left undecided "
            "because GitHub did not answer for any of them, so this reports no verdict"
        )
        return 71
    if unread:
        print(f"  NOTE: {unread} reference(s) were left undecided because GitHub did not answer")
    return 0


def utf8_console() -> None:
    """Write the console streams as UTF-8, whatever the host code page is.

    Reviewer-authored text carries characters such as emoji badges, which a Windows ANSI code
    page cannot encode. A stream without `reconfigure` is one a test harness substituted.
    """
    for stream in (sys.stdout, sys.stderr):
        reconfigure = getattr(stream, "reconfigure", None)
        if reconfigure is not None:
            reconfigure(encoding="utf-8", errors="replace")


def main(argv: list[str] | None = None) -> int:
    utf8_console()
    ap = argparse.ArgumentParser()
    ap.add_argument("cmd", choices=["attest", "claims", "comment", "status", "reply", "wait"])
    ap.add_argument("number", type=int)
    # No default, because the wrong repository is the failure this argument has actually had.
    # A default names one repository, and every run from elsewhere silently reads that one.
    # The number resolves there, the digest renders, and nothing in the output disagrees.
    # Two runs read a pull request here while one in their own repository was the subject.
    ap.add_argument(
        "--repo",
        required=True,
        metavar="OWNER/NAME",
        help="the repository the pull request is in, since a pull "
        "request number identifies no repository on its own",
    )
    ap.add_argument("--timeout", type=int, default=2700, help="seconds (default 45m)")
    ap.add_argument(
        "--pickup-grace",
        type=int,
        default=300,
        help="deprecated compatibility option, accepted but ignored",
    )
    ap.add_argument(
        "--check-grace",
        type=int,
        default=CHECK_GRACE,
        help="seconds a check may sit queued, or a required status go unposted, "
        f"before either reads as unstarted (default {CHECK_GRACE // 60}m)",
    )
    ap.add_argument(
        "--check-stall",
        type=int,
        default=CHECK_STALL,
        help="seconds a check may run before its duration is reported "
        f"(default {CHECK_STALL // 60}m)",
    )
    ap.add_argument(
        "--ignore-quota-signal",
        action="store_true",
        help="wait: request and poll the full --timeout even where the reviewer's own most "
        "recent activity elsewhere in this repository is a quota-limit refusal with nothing "
        "answering it since, or this pull request's newest Copilot review is a quota or "
        "error refusal, pass this once the quota is believed to have reset",
    )
    ap.add_argument(
        "--request",
        action="store_true",
        help="wait: request a Copilot round even on a fix push into a branch other than the "
        "default, which an attested local pass otherwise covers",
    )
    ap.add_argument(
        "--min-rounds",
        type=int,
        default=0,
        metavar="N",
        help="wait: treat the head as covered only once the reviewer's total rounds on this "
        "pull request, the same count `status` prints as `rounds=`, exceed N. Pass the "
        "`rounds=` a prior `status` already read, after re-requesting a review on a head that "
        "already carries one: `reviewed_head` alone would report that earlier round as done "
        "at once, never waiting for the fresh one the re-request was for (default 0, so any "
        "round counts, unchanged from before this flag existed)",
    )
    # `reply` takes the finding's words rather than its id.
    # There is deliberately no argument an id fits in, so the caller never holds one to mistype.
    ap.add_argument(
        "--match",
        metavar="TEXT",
        help="reply: words from the finding, matched against the thread's opening "
        "comment, and required to select exactly one unresolved thread",
    )
    ap.add_argument(
        "--path",
        metavar="FILE",
        help="reply: narrow --match to one file, for a file with several findings",
    )
    ap.add_argument(
        "--all-identical",
        action="store_true",
        help="reply: where --match selects several threads whose opening comments read the "
        "same, a finding posted twice, answer each of them with --body rather than refusing",
    )
    ap.add_argument(
        "--body",
        metavar="TEXT",
        help="comment or reply: the answer to post, carrying the fixing commit SHA or "
        "the evidence that disproves the finding",
    )
    ap.add_argument(
        "--resolve",
        action="store_true",
        help="reply: resolve the thread once the reply is confirmed",
    )
    ap.add_argument(
        "--checkout",
        metavar="DIR",
        help="attest: the checkout holding the pull request's head and its recorded local pass "
        "(default the current directory)",
    )
    a = ap.parse_args(argv)
    # Named for the command they belong to, since one silently ignored reads as one that took effect.
    # A `status` given --body reports a clean digest and writes nothing.
    # Nothing in that output says the reply never happened.
    reply_only = {
        "--match": a.match,
        "--path": a.path,
        "--resolve": a.resolve or None,
        "--all-identical": a.all_identical or None,
    }
    if a.cmd != "reply":
        for flag, value in reply_only.items():
            if value is not None:
                ap.error(f"{flag} belongs to `reply`, not `{a.cmd}`")
    if a.cmd != "wait" and a.request:
        ap.error(f"--request belongs to `wait`, not `{a.cmd}`")
    if a.cmd != "attest" and a.checkout is not None:
        ap.error(f"--checkout belongs to `attest`, not `{a.cmd}`")
    if a.cmd not in ("comment", "reply") and a.body is not None:
        ap.error(f"--body belongs to `comment` or `reply`, not `{a.cmd}`")
    required = ["--body"] + (["--match"] if a.cmd == "reply" else [])
    if a.cmd in ("comment", "reply"):
        values = {"--body": a.body, "--match": a.match}
        for flag in required:
            if not (values[flag] or "").strip():
                ap.error(
                    f"{a.cmd} requires a non-empty {flag}, since an empty answer records nothing"
                )
    if a.pickup_grace < 0:
        ap.error("--pickup-grace cannot be negative")
    if a.min_rounds < 0:
        ap.error("--min-rounds cannot be negative")
    # A negative threshold reports every check in that state, on every run, from the first read.
    # A field that fires always is one a reader learns to skip, which costs the real case.
    for name in ("check_grace", "check_stall"):
        if getattr(a, name) < 0:
            ap.error(f"--{name.replace('_', '-')} cannot be negative")
    # The two thresholds mean opposite things and the readings invert if the order does.
    # A stall under the grace reports a running check before it would report a starved one.
    # A case held the constants ordered while the flags could still be passed either way.
    if a.check_grace >= a.check_stall:
        ap.error(
            "--check-grace must be less than --check-stall, since a queued check is "
            f"judged sooner than a running one (got {a.check_grace} and {a.check_stall})"
        )
    # A bare name is the near-miss a required argument still admits, and unpacking it raises a
    # ValueError traceback rather than saying which half is missing.
    owner, _, repo = a.repo.partition("/")
    if not owner or not repo or "/" in repo:
        ap.error(f"--repo takes OWNER/NAME, not {a.repo!r}")

    if a.cmd == "claims":
        return check_claims(owner, repo, a.number)

    if a.cmd == "attest":
        return attest(owner, repo, a.number, a.checkout or ".")

    if a.cmd == "comment":
        return comment_on_pr(owner, repo, a.number, a.body)

    if a.cmd == "status":
        # One payload renders the digest and decides the code, for the reason `wait` reads one.
        # Fetched twice, a round landing between them prints one pull request and grades another.
        pr = full_read(owner, repo, a.number)
        out, _ = digest(owner, repo, a.number, pr=pr, grace=a.check_grace, stall=a.check_stall)
        print(out)
        return report_verdict(pr, owner, repo)

    if a.cmd == "reply":
        return reply_to_thread(
            owner, repo, a.number, a.match, a.body, a.path, a.resolve, a.all_identical
        )

    # `wait` mutates through the auto-request below, so it refuses a cross-owner target the way `comment` and `reply` do.
    # The policy is that a cross-owner target is not touched at all, so the refusal precedes the reading half too.
    ok, why = in_scope(owner)
    if not ok:
        print(f"status=OUT_OF_SCOPE nothing was written: {why}")
        return 64

    start = time.monotonic()
    pr = gql(Q_LIVE, owner, repo, a.number)
    done, answer, drift = liveness(pr, a.min_rounds)
    if done and a.ignore_quota_signal:
        full = gql(Q_FULL, owner, repo, a.number)
        if refusing_review(full) and not reviewed_head(full):
            a.min_rounds = max(a.min_rounds, len(reviewer_nodes(full, "reviews")))
            done = False
    # Read whenever nothing has landed on this pull request yet, whether or not a request is already outstanding.
    # An already-pending request drawing no answer at all is exactly the shape a repo-wide quota exhaustion leaves, measured over the six consecutive pull requests that followed a refusal and carried no Copilot activity at all.
    # The bot id for a fresh request comes from this same traversal, so a caller needing either pays for one call rather than two.
    history = [] if done or answer or drift else copilot_history(owner, repo)
    # `--ignore-quota-signal` only changes whether the signal below is acted on.
    # It does not change whether the history is read, since the auto-request line still needs it for the bot id regardless.
    # Read from the same, unfiltered history rather than one that drops this pull request's own entries: a genuine review on an earlier head of this same pull request, superseded since by a push, is real evidence about the account and not a self-reference to discard.
    # A refusal on this pull request's own current head still never reaches this signal, since it is caught directly and at higher priority first.
    signal = None if a.ignore_quota_signal else quota_signal(history)
    snapshot = None if done or drift else gql(Q_FULL, owner, repo, a.number)
    stopped = (
        None if a.ignore_quota_signal or answer or snapshot is None else stopping_refusal(snapshot)
    )
    held = (
        snapshot is not None
        and not a.request
        and not reviewer_requested(pr)
        and holds(snapshot)
        and (not answer or attested(snapshot))
    )
    # Request before the first poll, not just at the call site: a caller expects `wait` to make a review happen, not merely to watch for one.
    # Two prior gaps this closed, a push superseding an already-answered request and an auto-seed that never fired, both left nothing outstanding for the loop below to ever see land.
    # Skipped once a review already covers the head, once Copilot has already answered outside a formal review, or once something is already in the request set, so a second `wait` on the same PR never double-requests.
    recorded: bool | None = None
    final: dict | None = None
    if (
        not done
        and not answer
        and not drift
        and not signal
        and not stopped
        and not held
        and not reviewer_requested(pr)
    ):
        line, recorded = request_copilot_review(
            owner, repo, a.number, pr["id"], copilot_bot_id(history), POLL_DELAYS[0]
        )
        if recorded is False:
            final = gql(Q_FULL, owner, repo, a.number)
            if reviewer_requested(final):
                recorded, final = None, None
            else:
                line += ", and the request recorded nothing on the pull request"
        print(f"auto-request: {line}")
    if recorded is False:
        print(
            "note: the review request returned success and recorded nothing on this pull "
            "request, no pending reviewer and no review-request event, so this wait stops here "
            "rather than polling --timeout out against a request that does not exist."
        )
    elif held:
        print(
            "note: this pull request merges into a branch other than the default and Copilot "
            "has reviewed it already, so a fix push is covered by an attested local pass rather "
            "than another Copilot round, and this wait requests nothing. Pass --request to ask "
            "for a round anyway. An attested head has its checks polled instead, while they can "
            "still move its merge."
        )

        def still_held(polled: dict) -> bool:
            return (
                holds(polled)
                and local_cover(polled)
                and not unrecognized_shapes(polled)
                and held_checks_open(
                    polled,
                    check_nodes(polled),
                    datetime.now(UTC),
                    a.check_grace,
                    a.check_stall,
                )
            )

        assert snapshot is not None
        last_full = [snapshot]

        def held_read() -> dict:
            polled = gql(Q_HELD, owner, repo, a.number)
            if still_held(polled):
                return polled
            last_full[0] = gql(Q_FULL, owner, repo, a.number)
            return last_full[0]

        polled = backoff(last_full[0], held_read, still_held, start, a.timeout)
        final = polled if polled is last_full[0] else None
    elif stopped and not reviewer_requested(pr):
        print(
            "note: this pull request's newest Copilot review is a refusal naming the account "
            "quota or an error, which is what the weekly rate limit posts, so this wait requests "
            "nothing and stops here. Pass --ignore-quota-signal to request and poll anyway, "
            "once the limit is believed to have reset."
        )
    elif signal and not reviewer_requested(pr):
        # The poll below is skipped rather than shortened, because there is nothing partial about this signal.
        # The reviewer's own most recent word anywhere in the repository is the account quota, and nothing has answered it since.
        # Polling this pull request's own silence for up to 45 minutes would only relearn that same account state a call late.
        number, hist_refusal = signal
        print(
            f"note: the reviewer's own most recent activity anywhere in this repository, on "
            f"pull request #{number} at {hist_refusal.get('_at') or 'an unknown time'}, is a "
            "quota-limit refusal with nothing answering it since, so this wait stops here "
            "rather than polling --timeout out against the same account state. Pass "
            "--ignore-quota-signal to poll anyway, once the quota is believed to have reset."
        )
    else:
        backoff(
            (done, answer, drift),
            lambda: live_state(owner, repo, a.number, a.min_rounds),
            lambda s: not any(s),
            start,
            a.timeout,
        )

    # One payload decides the digest and the exit code together.
    # Read separately, a review landing between them prints coverage and returns a timeout code.
    # A reader resolves that by believing the code, dropping the review it was just shown.
    # The digest also earns its call at the timeout.
    # A bare PENDING line reports a broken wait and a slow reviewer identically.
    final = full_read(owner, repo, a.number, final)
    now = datetime.now(UTC)
    # Parsed here and handed down, so the digest and the exit code share one read of the rollup.
    # Deriving the stuck shapes from that list costs no parse, which is what was doubled.
    checks = check_nodes(final)
    stuck = checks_stuck(checks, now, a.check_grace, a.check_stall)
    out, _ = digest(
        owner,
        repo,
        a.number,
        pr=final,
        now=now,
        grace=a.check_grace,
        stall=a.check_stall,
        checks=checks,
    )
    print(out)
    print(f"waited={int(time.monotonic() - start)}s")
    # The shape reading comes first and is not gated on coverage of the head.
    # `reviewed_head` is itself one of the readings a drift breaks.
    # A renamed login matches no filter here, so it reads as no review at all.
    # Every arm below then reports a review that landed as a pending one.
    # Gating the verdict behind it left the login check unable to reach an exit code.
    # The digest above printed `shapes=UNRECOGNIZED` the whole time it did so.
    # Coverage of the head is the other half, returning 0 only once the diff is covered too.
    held = held and holds(final)
    covered = held and local_cover(final)
    halted = None if a.ignore_quota_signal else stopping_refusal(final)
    if (
        unrecognized_shapes(final)
        or covered
        or (not halted and head_review_done(final, a.min_rounds))
    ):
        verdict = report_verdict(final, owner, repo)
        # The check reading ranks under both of those, and never replaces either.
        # An unreadable shape means no field here can be believed, this one included.
        # A partial round means the diff is part-reviewed, which outranks a wedged gate.
        # Only once the review itself is sound does a stuck required check decide the code.
        if verdict:
            return verdict
        if covered and held_checks_open(final, checks, now, a.check_grace, a.check_stall):
            print(
                "status=CHECKS_PENDING an attested local pass covers this head, and by the "
                "timeout a check had not concluded, none had posted, or the merge was still "
                "UNKNOWN, so the merge is not ready yet: wait again, or read the checks above"
            )
            return 30
        # The review loop closing is not the merge gate, and 0 alone was saying it was.
        # A wait ends the moment coverage lands, which leaves the checks mid-flight nearly always.
        # So a merely pending check is not this code, or the code would be the usual outcome.
        # Only a shape no waiting clears earns it, which is what the stuck field already prints.
        # It is read from the same payload the digest was, so the two can never disagree.
        # A rollup carries checks the ruleset does not require, four of six on a green run here.
        # So only a stuck check whose rollup node reads `isRequired` counts toward the code.
        # Without that, a stuck optional check on a merge a thread blocks returns 44.
        # `BLOCKED` is required as well, since `CLEAN` proves no required gate is outstanding.
        # The digest reports every stuck check either way, so the narrower code costs nothing.
        gates = [n.get("name") or "unnamed" for n, _ in stuck if n.get("required")]
        if final.get("mergeStateStatus") == "BLOCKED" and gates:
            # The block above prints optional stuck checks too, so the required ones are named.
            # They need not be all that blocks it.
            # `BLOCKED` is also worn by an open thread or a missing approval.
            print(
                "status=CHECKS_NOT_MERGEABLE the review loop is closed, the merge reads "
                "BLOCKED, and a required check is in a shape waiting does not clear, "
                f"required stuck {', '.join(repr(g) for g in gates)}: read their lines in the "
                "block above, since a starved check wants a re-run, an unposted one its poster, "
                "a long one a judgment, and a failed one a fix. It need not be all that blocks "
                "the merge, because BLOCKED is also worn by a thread or a missing approval"
            )
            return 44
        return 0
    # A refusal before an answer, since it names the round that declined where 40 names none.
    # The digest prints both bodies regardless, so the narrower code costs the reader nothing.
    refusal = refusing_review(final)
    if refusal is None and not a.ignore_quota_signal:
        refusal = stopping_refusal(final)
    if refusal and quota_refusal(refusal):
        print(
            "status=COPILOT_QUOTA_EXHAUSTED the newest Copilot review declined because the "
            "requesting account has reached its Copilot review quota, printed above under "
            "COPILOT REFUSED THIS ROUND: that is an account-level state, not one this pull "
            "request or a re-request clears, so proceed on the coverage the other reviewers "
            "already gave this pull request rather than waiting on Copilot again"
        )
        return 46
    if refusal and (limit := confirmed_quota(owner, repo, refusal)):
        print(
            "status=COPILOT_QUOTA_EXHAUSTED the newest Copilot review on this pull request says "
            "it encountered an error, and the reviewer's own run log names the cause: "
            f"{limit}. Do not re-request until the limit resets, then pass "
            "--ignore-quota-signal, and proceed meanwhile on the coverage the other reviewers "
            "already gave"
        )
        return 46
    if refusal and possible_quota(refusal):
        print(
            "status=COPILOT_ERROR_POSSIBLE_QUOTA the newest Copilot review on this pull request "
            "says it encountered an error and did not review, printed above under COPILOT "
            "REFUSED THIS ROUND, which is the body the weekly rate limit posts. Read it as a "
            "possible quota hit: do not re-request, proceed on the coverage the other reviewers "
            "already gave, and hand the state to the maintainer, who can pass "
            "--ignore-quota-signal once the limit is believed to have reset"
        )
        return 46
    if held and not refusal:
        print(
            "status=AWAITING_LOCAL_PASS no Copilot round covers this head and none was requested, "
            "since a fix push after the first round is covered by a local pass. Run the local "
            "strict review, record it, push, and run `pr_review.py attest` from the checkout, "
            "or pass --request to ask Copilot for a round instead"
        )
        return 49
    if refusal:
        print(
            "status=REVIEW_IS_A_REFUSAL the review carrying the head says it did not review, "
            "so it covers nothing and no further review follows it: read the body above, "
            "since a file-count refusal is cleared by splitting the pull request, and "
            "re-requesting this head clears nothing"
        )
        return 41
    # An answer before a stall, because the reviewer saying something outranks it saying nothing.
    if answered_outside_review(final):
        print(
            "status=ANSWERED_OUTSIDE_REVIEW the reviewer answered without reviewing, "
            "so read the comment above and decide, since where it declines or names a limit "
            "no review follows and re-requesting does not clear it"
        )
        return 40
    if recorded is False:
        print(
            "status=REQUEST_NOT_RECORDED the Copilot review request returned success and "
            "added neither a pending reviewer nor a review-request event to this pull request. "
            "The likely cause is an exhausted Copilot allowance on the requesting account, "
            "which is the maintainer's to restore: clearing the request set and requesting "
            "again does not clear it, so hand it to the maintainer"
        )
        return 48
    # Lowest priority of the terminal readings, since it is inferred from elsewhere in the repository rather than read on this pull request directly.
    if signal:
        number, hist_refusal = signal
        print(
            "status=COPILOT_QUOTA_EXHAUSTED_REPO_WIDE this pull request's head carries no "
            f"Copilot review or comment of its own, and the reviewer's own most recent activity "
            f"in this repository, on pull request #{number} at "
            f"{hist_refusal.get('_at') or 'an unknown time'}, was a refusal citing the "
            "account quota limit rather than a review of this head. Proceed on the coverage "
            "the other reviewers already gave this pull request, or pass --ignore-quota-signal "
            "to poll this pull request's own head for the full --timeout"
        )
        return 47
    print("status=PENDING")
    return 30


if __name__ == "__main__":
    sys.exit(main())