fix(cb-r9ac): the unchanged-work guard measures a start against the queue the pass found

[rmstdope]

Summary

• Bead cb-r9ac. start_due recorded a role's fingerprint after facts.take had removed the handed bead, so a pass that did exactly its one bead left the live queue equal to the record and the unchanged-work guard held the next start. In atlantis-hud the only bugfixer sat idle 20+ minutes with three candidates queued.
• The fingerprint is now taken before the hand-off. A pass that closes its bead reads as changed and starts again; a pass that gives its bead back untouched reads as unchanged and is held, which is the guard's purpose.
• New start_due test: one producer, two beads, one handed; after the pass ends, the guard does not hold with the bead gone and does hold with the queue as found.
• Also rewrites the one grep -q pipeline in tests/worktree-safety.sh (from fix(cb-7suc): the worktree pruner removes finished trees in a repo with submodules or squash merges #471) that scripts/pipefail-pipelines refuses, so the local gate is green again. CI does not run that check, which is why fix(cb-7suc): the worktree pruner removes finished trees in a repo with submodules or squash merges #471 passed.

Test plan

• cargo test --locked --bin cerebro-tui a_pass_that_took_its_bead_off_the_queue fails before the fix, passes after
• bash tests/gate green

🤖 Generated with Claude Code

https://claude.ai/code/session_01GKiLr59caM9f6DxkrsDYR8

[Copilot]

Copilot review overview

🔵 Needs a closer look

The change alters load-bearing fleet-supervision timing logic whose correctness depends on subtle within-tick ordering, warranting a final human check.

Review effort: Balanced
Findings: None

What changed in this PR

This PR fixes a timing bug in the fleet view's start_due (fleet-view/src/main.rs), where the "unchanged-work" guard's fingerprint was recorded after facts.take had already removed the just-handed bead from the queue. As a result, a pass that closed exactly its one handed bead left the live queue equal to the recorded fingerprint, so the guard misread it as "nothing changed" and held the next start — in atlantis-hud the only bugfixer sat idle for 20+ minutes with three candidates queued. The fingerprint is now captured before the hand-off, so a pass that closes its bead reads as changed (starts again) while a pass that returns its bead untouched reads as unchanged (correctly held).

Changes:

• Capture the start fingerprint (found) before facts.take in start_due, and use it in ledger.note_started.
• Add explanatory doc comments in triggers.rs::unchanged and main.rs describing the "queue as found" semantics.
• Add a regression test (a_pass_that_took_its_bead_off_the_queue_is_not_held_as_unchanged) covering both the closed-bead and returned-bead cases.
• Rewrite the grep -q pipeline in tests/worktree-safety.sh that scripts/pipefail-pipelines refuses, keeping the local gate green.

File	Description
fleet-view/​src/​main.rs	Records the start fingerprint before facts.take; adds a regression test for the closed vs. returned-bead cases.
fleet-view/​src/​triggers.rs	Adds a doc comment clarifying the guard measures against the queue the pass found.
tests/​worktree-safety.sh	Replaces the pipefail-unsafe grep -q pipeline with an equivalent bash glob check on the porcelain output.

I verified the fix: found is computed before facts.take(id) and passed to note_started; the new test genuinely fails before the fix (the recorded fingerprint would drop the handed bead) and passes after. The worktree-safety.sh rewrite preserves the original end-of-line anchoring semantics (the appended echo restores a trailing newline stripped by command substitution, and the quoted "worktree $tree" is matched literally). Helper functions and public symbols used by the test all exist and match.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.