Rename a loop when its session is renamed (#261)

[scgopi]

Fixes #261.

Renaming a session from inside it (/rename hello) left the loop's card showing whatever title it was created with. The title is the one field a loop is expected to outgrow — it is written before the work exists — so the two names disagreed at exactly the moment the human had just said what the loop is.

Where the answer comes from

Claude Code writes it down, in the transcript GraphCode already knows how to find:

{"type":"custom-title","customTitle":"hello","sessionId":"a825abc7-…"}

It is a standalone record, re-emitted at every checkpoint rather than once at the rename, so a tail read always carries the current one and the last one wins. Read on the existing 15s presence poll, applied through the same renameNode a human's rename goes through — so a blank title is refused and everything a rename deliberately leaves alone (id, session, edges, layout) stays left alone.

Three things this is careful about

custom-title only. Every session also carries an agent-name record, and it is not the same thing: it holds a name the CLI assigned itself (angleReuse2) and is present in sessions nobody has renamed. Reading it would retitle most of a graph to strings no human ever typed — a far worse bug than the one being fixed. anAgentNameIsNotARename pins this.

Once per change, not once per sight. LoopNode.sessionTitle records what the session was last observed to be called, which is not what the node is called — the two part company the moment somebody renames the card from the sidebar, and that gap is what keeps the card's name theirs instead of having it overwritten on the next tick. It is persisted for the same reason: a daemon restart must not make an applied rename look new again.

Cost. summarisesLoops is off by default, so on a default install nothing was reading transcripts at all; ungated, this would have added a 512KB tail read per loop every fifteen seconds to every install there is, to re-read a name that changes when a human types six characters. Gated on TranscriptFreshness — its own instance, not shared, because hasChanged is consuming and the two readers would otherwise take turns to work — a quiet loop costs one stat.

Deliberately out of scope

• Remote loops. Reading one costs an ssh probe per node per poll. Worth it for a rail that moves every few seconds, not for a name that changes once in a loop's life. Reported as "nothing said" rather than bought at that price.
• The other three backends. None of Copilot CLI, Codex, or OpenCode has a rename that leaves a record behind, so they return nil rather than a guess.
• The reverse direction. An app-side rename is not pushed into the session.

Regression surface

Risk	Guard
Clobbering a human's title every 15s	sessionTitle change-detection — aRenameFromTheSidebarSurvivesTheNextPoll
Renaming loops from agent-name	Record type matched exactly — anAgentNameIsNotARename
The poll starting to write graph.json every tick	Persist only when a rename actually landed; the existing thePollMustNotPersistTheGraph still passes
Starving the summary rail's freshness gate	Separate TranscriptFreshness instance; SummaryRailTests unchanged and green
Old graphs failing to decode	decodeIfPresent, matching every other field on LoopNode

Verification

Check	Result
xcodebuild test	✅ 1534 tests / 159 suites, up 11
graphcode / graphcoded / graphcode-cli builds	✅ all three succeed
swiftlint lint	✅ 0 errors
swift format lint --strict	✅ clean

🤖 Generated with Claude Code

https://claude.ai/code/session_011hEyiK8W9XW4Tq6je98kTX

[scgopi]

Don't merge the PR yet, need to see if it is low risk