routine-self-review: cap consecutive CONSOLIDATE nights (closes #203)

docs/dream-cycle/2026-11-05-routine-self-review-report.md

@@ -0,0 +1,256 @@
+# Dream Machine Self-Review — 2026-11-05
+
+## TL;DR
+Last week's forecast (PR #202, 2026-10-29) scores a genuine **HIT**: every
+scheduled night since then left a live GitHub trace, unlike the vacuous HIT
+scored the week before. But the six-night CONSOLIDATE run it measured points
+at the real structural gap: `main` has not advanced since 2026-10-19 (PR
+#196), 21 `dream/*` candidate PRs sit open against
+`implementationTrack.reviewBacklogLimit=6`, and because at least one eligible
+draft always exists, STEP 1.5's MODE has resolved to CONSOLIDATE on every
+single non-self-review night since — which explicitly skips STEPS 2–9 and
+17–24. There is no upper bound on how long this can run: CONSOLIDATE's entry
+condition only clears once a human merges enough PRs to drop the backlog
+below the limit, something this routine can never do by its own merge policy.
+Tonight's candidate adds a `CONSOLIDATE_STREAK` cap (7 non-self-review nights)
+so the routine is forced back to RESEARCH-ONLY — and keeps producing new
+findings — even while the review backlog never clears.
+
+## Score: last self-review's forecast (PR #202, issue #201)
+**Forecast (frozen 2026-10-29):** over the next 7 nights, every scheduled
+firing of this routine leaves at least one GitHub-visible trace (a
+`dream/<DATE>-*` branch, or an equivalent reachable record), i.e. zero further
+nights with no trace at all. Current value at freeze time: 6 of the prior 7
+nights (2026-10-23→10-28) left no trace. Target: 0 of the next 7.
+
+**Measured (2026-10-30 → 2026-11-04, live GitHub check — `search_issues`,
+`pull_request_read.get_comments` on PRs #161/#180/#182/#184/#189/#192):**
+- 2026-10-30: PR #161 — CONSOLIDATE comment, `<!-- dream-consolidate:
+  3fc3222d -->` marker, merge commit `4af4332`.
+- 2026-10-31: PR #180 — CONSOLIDATE comment, merge commit `63f51fc`.
+- 2026-11-01: PR #182 — CONSOLIDATE comment, merge commit `1d563fae`.
+- 2026-11-02: PR #184 — CONSOLIDATE comment, merge commit `f51e51ac`.
+- 2026-11-03: PR #189 — CI-diagnostic + main-merge-in commit `7e19136`
+  (consolidate marker posted retroactively the following night, same action).
+- 2026-11-04: PR #189 marker comment + PR #192 — full CONSOLIDATE comment,
+  merge commit `7104179`.
+- No `dream/2026-11-0{1,2,3,4,5}-*` or `dream/2026-10-3{0,1}-*` branch exists
+  anywhere (`list_branches` confirmed) — consistent with CONSOLIDATE opening
+  no new candidate.
+- No new `[Dream Cycle ...]` issue was filed in the window (`search_issues
+  is:issue created:2026-10-30..2026-11-05` → 0 results).
+
+**Verdict: HIT, and not vacuous this time.** Six real nights ran, each
+left a reproducible GitHub trace, and the guard's own premise — "use
+whichever path is reachable" — was exercised: every comment above was
+posted via the GitHub MCP tool while `gh auth status` was separately
+observed broken in this session (`GH_TOKEN` invalid), same as the two
+preceding self-reviews. The forecast's mechanism held.
+
+## Measure
+- `search_issues is:pr is:merged head:dream/`: 39 merged. `is:pr is:unmerged
+  is:closed head:dream/`: 7 closed-unmerged. `is:pr is:open head:dream/`: 21
+  open. `pipelineMergeRate = 39 / (39+7+21) = 58.2%`.
+- `reviewBacklogSize = 21` (open dream/* candidate PRs), `6×` over
+  `implementationTrack.reviewBacklogLimit=6`.
+- `ledgerStale=true` — main's `LEDGER.md` last real (merged) row is
+  2026-10-12 (#177); every self-review/CONSOLIDATE row since lives only on
+  unmerged branches.
+- No REJECT, INCONCLUSIVE, closed-unmerged or stale candidate exists *since*
+  the last self-review (2026-10-29) to classify by cause — because no new
+  candidate was proposed at all in that window. That absence is itself the
+  measurement.
+- Confirmed via `list_branches`/`search_pull_requests`: the last *new*
+  `dream/<DATE>-<surface>` branch (i.e. a non-self-review RESEARCH/IMPLEMENT/
+  RESEARCH-ONLY night) is `dream/2026-10-18-security-adversarial` (PR #194).
+  Every one of the 16 non-self-review nights since (2026-10-19→2026-11-04,
+  excluding the self-review nights 10-22/10-29) resolved to CONSOLIDATE.
+
+## Cluster
+Taxonomy: planning | tool | memory | verification | recovery | resource |
+security | unknown. With zero new candidates in the window there is nothing
+to classify by individual failure cause (same shape as last week). The
+dominant and only cluster this time is **planning**: STEP 1.5's own MODE
+state machine has no exit from CONSOLIDATE once the backlog crosses
+`reviewBacklogLimit` and at least one eligible draft exists — which, because
+a new "oldest unmarked" draft is always available among 21 open PRs, holds
+on every night regardless of how many nights pass. The gate that could clear
+it (the backlog dropping below the limit) is entirely human-controlled; this
+session's own merge policy forbids it from closing that gate itself. This is
+a defect in the decision rule's shape, not in any evaluator, memory,
+recovery path, resource limit or security control — hence **planning**,
+count 1 (dominant by construction: it is the only mechanism active across
+all 16 affected nights).
+
+## Candidate (mechanism, one sentence)
+Cap CONSOLIDATE at 7 consecutive non-self-review nights (`CONSOLIDATE_STREAK
+< CONSOLIDATE_STREAK_CAP`, new constant = 7 in `packages/compile/src/v2.ts`);
+once reached, MODE forces RESEARCH-ONLY instead, even with an eligible draft
+still waiting, so new findings keep arriving regardless of how long the
+human review backlog takes to clear.
+
+`packages/compile/src/v2.ts`:
+- Added `export const CONSOLIDATE_STREAK_CAP = 7`.
+- `v2ModeSelection()`: added a `CONSOLIDATE_STREAK` measurement instruction
+  (consecutive non-`routine-self-review` nights with neither a new
+  `<branchPrefix><DATE>-<surface>` branch nor a new `[Dream Cycle <DATE>]`
+  issue — i.e. nights whose only visible action was CONSOLIDATE); added
+  `AND CONSOLIDATE_STREAK < 7` to CONSOLIDATE's entry condition; added
+  `OR CONSOLIDATE_STREAK >= 7` to RESEARCH-ONLY's entry condition; added
+  `CONSOLIDATE_STREAK` to the printed MODE line.
+
+7 was chosen to match the one full week already observed running
+uninterrupted (2026-10-30→2026-11-04 plus the two flanking non-self-review
+days) — it guarantees at least one research-producing night per
+self-review period, without requiring a new config field or touching
+`implementationTrack.reviewBacklogLimit`/`selfReview.everyDays` (both
+explicitly off-limits to self-review). Does not touch any item on the
+routine constitution: `reviewBacklogLimit` and `selfReview.everyDays` are
+unchanged values; CONSOLIDATE's never-merge/never-close/never-force-push
+limits, the independent critic/reward-hack checks, evaluator entrypoints,
+witness computation, ledger schema, research floors, and config
+validation/v1 golden snapshots/`v2.test.ts` are all untouched in substance —
+confirmed by diff review.
+
+## Evaluation Receipt
+Real evaluator: `npm test` (`vitest run && node --test scripts/*.test.mjs`).
+
+| | Baseline (`3fc3222d`, parent, post-unshallow) | Candidate |
+|---|---|---|
+| vitest | 886 | 887 (+1 new, 0 regressions) |
+| governance | 195 | 195 (unchanged) |
+| lint / typecheck | clean | clean |
+| mirror check (`compile-targets.mjs --check`) | pass | pass (all 5 targets regenerated + reviewed) |
+
+Self-hosted `dream.config.json` compile output confirmed byte-identical to
+its regenerated mirror (`docs/dream-cycle/compiled/ruvnet-dream-machine.md`);
+all 5 targets sharing this compile engine (self-hosted +
+`ruflo`/`metaharness`/`ruvector`/`ruclip`) had their mirrors regenerated —
+same clause shape, verbatim across targets since neither touched function
+takes a per-target parameter.
+
+## Baseline
+Parent commit `3fc3222d2e5566a203331d3686c4d2eaf06713a8` (`main`), this
+session's own checkout (unshallowed from the fresh clone's default
+`clone_depth` cutoff to resolve the same already-tracked environmental
+governance-test failure as PR #202/#150 — `BASELINE_COMMIT`-unreachable
+`packages/edge-contracts/src/index.ts`; not a code change).
+
+## Darwin Lineage
+`DARWIN=not-applicable` — a Schedule-contract/MODE-selection prose-and-
+constant fix has no evolvable population or fitness function.
+
+## Evidence
+OBSERVATION: `main` unchanged at `3fc3222d` since PR #196 (2026-10-19).
+OBSERVATION: 21 open `dream/*` PRs, 0 new ones since PR #194 (2026-10-18).
+OBSERVATION: 6 consecutive non-self-review nights (2026-10-30→11-04), each
+independently confirmed via PR comments, all CONSOLIDATE, zero new branches
+or issues. INFERENCE: this pattern can repeat indefinitely because
+CONSOLIDATE's entry condition does not depend on elapsed time or streak
+length, only on backlog size and eligibility, both of which a 21-PR, 0-merge
+backlog keeps satisfied night after night. DECISION: add a streak cap rather
+than touch the backlog limit itself, since the limit is off-limits to
+self-review and is not, on its own, the mechanism causing the lock — the
+*uncapped duration* of CONSOLIDATE is.
+
+## Reward-Hack Check
+Independent critic dispatched (separate-context subagent, no authorship
+stake, given only the diff, `ROUTINE_CONSTITUTION`, and instructions to
+attack it against the off-limits list, test vacuity, and specification
+ambiguity): **CLEAR**. It independently re-ran `npm run build -w
+packages/compile`, `vitest run` on the three affected test files (181
+passed), `lint`, `typecheck`, and `compile-targets.mjs --check` (all green),
+confirmed the `CONSOLIDATE_STREAK` "new branch OR new issue" discriminator
+is exhaustive against the actual STEP numbering (RESEARCH-ONLY always files
+an issue per STEPS 15–18; IMPLEMENT/RESEARCH file both; only CONSOLIDATE
+does neither), and confirmed the new test is not vacuous (asserts exact
+substrings absent pre-change). It flagged two non-blocking nitpicks: the
+live-check command given (`gh pr list --state open ...`) is open-PR-only, so
+a merged-and-branch-deleted night's branch signal could be missed — though
+since every non-CONSOLIDATE night also files a persistent issue, this
+under-detection only biases toward *more* research, not toward dodging the
+cap, so it is not mission-defeating; and the cap value (7) is a defensible
+but somewhat round choice, with a smaller cap arguable as more conservative.
+Both disclosed here rather than silently accepted. Self-critique in
+parallel: does this
+candidate weaken a test, benchmark, threshold, or gold answer? No — one test
+added (`v2.test.ts`), none removed or loosened, and the new assertions check
+both the presence of the new clauses and that the raw constant name
+(`CONSOLIDATE_STREAK_CAP`) never leaks into the compiled prompt (only its
+numeric value should). Does it cherry-pick or exploit the evaluator? No —
+`npm test` runs unmodified. Does it change `reviewBacklogLimit` or
+`selfReview.everyDays`? No — both are read, not written, same as before.
+Does forcing RESEARCH-ONLY after the cap let a future CONSOLIDATE night
+dodge the cap by, say, posting a spurious issue to reset the streak without
+doing real research? The streak's reset condition is tied to STEPS 2-9/15-18
+actually running (a RESEARCH-ONLY night necessarily files a real `[Dream
+Cycle <DATE>]` issue with real research content per STEP 17-18's own
+requirements, which are untouched) — not independently satisfiable without
+doing the work the cap is meant to compel.
+
+## Security Review
+Touches only compiled instructions for a future agent run — no new tool,
+permission, credential, or network call. It reduces a resource/availability
+risk (new research silently withheld for unbounded nights while backlog
+accumulates) rather than introducing one. No prompt-injection surface, no
+new authority granted to any future night's agent beyond *when* it may defer
+research in favor of CONSOLIDATE.
+
+## Regression Analysis
+10 files changed: 2 logic files (`v2.ts`, `index.ts`, one new constant + one
+re-export + the STEP 1.5 template text), 1 test file (+1 test block), 2
+generated snapshot files, 5 mechanically-regenerated compiled mirrors. All
+887 vitest + 195 governance tests pass; lint/typecheck clean; mirror check
+clean. No GLOBAL INVARIANT, verdict rule, promotion gate, merge policy
+(`autoMerge`/`prMode`/ready-for-review predicate), CONSOLIDATE's never-merge/
+close/force-push limits, independent-critic/reward-hack check, evaluator
+entrypoint, witness computation, ledger schema,
+`reviewBacklogLimit`/`everyDays` values, research floor, or
+config-validation/golden-snapshot test was touched in substance — all
+required by the self-review constitution (ADR-0011) to stay out of bounds,
+confirmed by diff review.
+
+## ADR
+None — a MODE-selection threshold/prose fix to the routine's own compiled
+template, not an architectural decision (STEP 19: never for parameter
+changes or minor prompt/routing tweaks).
+
+## Gist
+`GIST=LOCAL` — no `gh` CLI (`GH_TOKEN` invalid in this session) and no
+gist-creation MCP tool available tonight; same condition as PR #199/#202.
+Full report committed as
+`docs/dream-cycle/2026-11-05-routine-self-review-report.md`, the durable
+artifact.
+
+## Issue
+#203
+
+## Witness
+
+```
+report_sha256 : dfe9cb8068f2b6d469c7b6d8f98363a16dee565452cd11fd2891dd052cc9f549
+session_commit: 3fc3222d2e5566a203331d3686c4d2eaf06713a8
+witness       : d67dfbb0019792117619684ca623ce88dff9c60a8db7e9cd6adf5264c73d8afc
+```
+
+Computed over this file's canonical bytes (everything above this heading,
+trailing whitespace collapsed to one newline). Verify:
+
+```bash
+dream-machine witness verify-report <committed-report-file>
+```
+
+## Merge Policy
+**This PR stays draft, regardless of verdict.** Per STEP 1.6/ADR-0011's
+routine constitution: "A self-review PR always stays human-review-only,
+whatever `prMode` says." This session never merges and never autonomously
+promotes candidate state.
+
+## Ledger
+Row to be appended for 2026-11-05 (`docs/dream-cycle/LEDGER.md`). Forecast
+frozen for the next self-review (2026-11-12, `EPOCHDAY % 7 == 0`): over the
+next 7 nights, at least one of them resolves to RESEARCH-ONLY, IMPLEMENT or
+RESEARCH (i.e. `CONSOLIDATE_STREAK` is observed to reset at least once,
+proving the cap fires in practice and not just in the template text).
+Current value: 0 resets observed in the prior 16 non-self-review nights
+(2026-10-19→2026-11-04, all CONSOLIDATE). Target: ≥1 of the next 7.

docs/dream-cycle/LEDGER.md

@@ -53,3 +53,4 @@
 | 2026-10-10 | compiler-parity | slots[].deep/scan[]/bonusModuli/cron/evaluatorEntrypoints/labels/extraDisciplines/competitors/branchPrefix/ledgerPath/adrConvention.dir values could break out of the compiled prompt's fences/inline-spans, injecting fabricated instructions into this self-hosting system's own authoritative control prompt; named-but-unclaimed since PR #143; closed across the full class via 3 independent-critic rounds (BLOCKED, BLOCKED, CLEAR) plus 1 self-found packageSpec-escaping gap | #164 | #165 | yes | ACCEPT | npm test 769->795 vitest (+26), 159/159 governance unchanged, 0 regressions; lint/typecheck clean; self-hosted dream.config.json compile output byte-identical throughout | faed331494a93b8d038533828f8402a4a8d5eac0b788cf60c916f3cb49191a8c | 15 open dream-cycle/maintenance PRs (#98,#101,#137,#139,#141,#145,#148,#150,#152,#154,#156,#158,#159,#161,#163), 0 overlapping this candidate; LEDGER.md 10d stale on main before this row (ledger-signals-surface finding, out of scope for compiler-parity slot); no session merge or self-promotion |
 | 2026-10-11 | ledger-signals | duplicateDirections counted raw row occurrences instead of distinct nights; a single re-appended night (2026-08-28) false-positived as a 3-night repeat; dedupe by (directionKey, date) | #166 | #167 | yes | ACCEPT | npm test 928->929 (+1), 0 regressions; live: real ledger's 2026-08-28 duplicateDirections false positive gone post-fix | 54824cd4528bcc15932c98bd5d97dd2222271f571a1f7fd0ce32a29ddf790401 | zeroMergeStreak=false (36 dream/* PRs confirmed merged via search_pull_requests ground truth, last #143 on 2026-09-30); 14 open unmerged dream/* PRs since then (#98,#101,#137,#139,#141,#145,#148,#150,#152,#154,#156,#158,#161,#165), none overlapping this candidate's file; ledgerStale=true (11d) on main before this row; issue #149/PR #150 (shallow-clone git-show oracle) found already filed, withheld not re-fixed; no session merge or self-promotion |
 | 2026-10-12 | evaluation-adapters | evaluateHarnessCandidate's hardReject keyword list omitted 'no declared harness surface'; a zero-surface patch lacking holdout evidence was misclassified INCONCLUSIVE instead of REJECT (structural invalidity masked as insufficient evidence); added the missing keyword | #176 | #177 | yes | ACCEPT | npm test 1001->1004 (+3 vitest, 0 regressions); live repro: zero-surface+no-holdout candidate flips INCONCLUSIVE->REJECT, same score/reasons; lint/typecheck clean | 6fb594bd01e09a8ed25b67833f50601a8b27c772bc202ccce4c84f48b74af9e4 | zeroMergeStreak=false via ground-truth search_pull_requests (38 dream/* PRs merged, most recent #167 on 2026-10-11); #137/#148/#156 (same evaluation-adapters surface) remain open/draft/unreviewed, read in full, no overlap (different files/functions); harness-optimizer.ts (PR #30) has zero call sites anywhere in the live pipeline, flagged as a wiring gap, not fixed tonight; no session merge or self-promotion |
+| 2026-11-05 | routine-self-review | PR #202's forecast scored HIT (non-vacuous: 6 real CONSOLIDATE nights, each left a GitHub trace), but that same run exposed the real gap: CONSOLIDATE has no exit while 21 open dream/* PRs stay over reviewBacklogLimit=6 and an eligible draft always exists, so main has gone 16 non-self-review nights (since PR #194, 2026-10-18) with zero new research; added CONSOLIDATE_STREAK cap (7 consecutive non-self-review nights) forcing RESEARCH-ONLY past the cap | #203 | #204 | yes | ACCEPT | npm test 886->887 vitest (+1 new, 0 regressions), 195/195 governance unchanged; lint/typecheck/mirror-check clean; self-hosted + 4 other targets' compiled mirrors regenerated; CONSOLIDATE_STREAK cap=7 added to STEP 1.5; independent critic subagent CLEAR | d67dfbb0019792117619684ca623ce88dff9c60a8db7e9cd6adf5264c73d8afc | PR #202 (last self-review, 2026-10-29) stays draft/unreviewed; 21 open dream/* PRs, 0 new since PR #194 (2026-10-18); 6 confirmed CONSOLIDATE nights 2026-10-30->11-04 on PRs #161/#180/#182/#184/#189/#192; zeroMergeStreak=false (39 merged historically, most recent #177 2026-10-12); reviewBacklogSize=21; pipelineMergeRate=58.2%; no session merge or self-promotion |