test(strategy): prove dangling checkpoint trailer for IDLE + empty FilesTouched

[Soph]

What

Adds a RED regression test that reproduces and proves a silent prepare-commit-msg / post-commit mismatch: a commit can end up carrying an Entire-Checkpoint trailer that points at a checkpoint that was never materialized locally (a dangling trailer).

This is a test-only, TDD-first PR — the test is intentionally failing and documents the invariant a follow-up fix must restore. CI will be red until the fix lands.

New file: cmd/entire/cli/strategy/phase_postcommit_dangling_trailer_test.go → TestPostCommit_IdleSession_NewLiveTranscriptEdits_DanglingTrailer.

Scenario (reconstructed from a real Pi session)

1. A session is condensed (clears state.FilesTouched, stamps LastCheckpointID), then goes IDLE.
2. A new turn arrives without a TurnStart (rapidly submitted prompt), so the session stays IDLE and SaveStep never repopulates FilesTouched. No shadow branch exists for the new edits yet.
3. The agent edits files and commits them.

Root cause

• PrepareCommitMsg re-extracts the edited files from the live transcript (sessionHasNewContentFromLiveTranscript), overlaps them with the staged files, and adds the trailer.
• PostCommit computes filesTouchedBefore only from persisted state.FilesTouched for non-ACTIVE sessions (the active-only fallback in manual_commit_hooks.go:1246-1252) and never passes the committed files as an overlap signal. With FilesTouched empty and no staged files, hasNew collapses to false and/or the guard in shouldCondenseWithOverlapCheck (len(filesTouchedBefore)==0 → return false) rejects condensation.

Net: the commit keeps the trailer, but no checkpoint is ever written for it.

What the test proves

• Premise passes — PrepareCommitMsg adds the trailer purely from the live transcript (log: prepare-commit-msg: trailer added ... checkpoint_id=...).
• Invariant fails — after commit + PostCommit, entire/checkpoints/v1 does not exist and LastCheckpointID stays pinned to the prior checkpoint. The trailer dangles.

Notes for the fix (out of scope here)

• TestPostCommit_IdleSession_NoTranscriptFallbackForCarryForward explicitly asserts IDLE sessions must NOT re-extract from the transcript (it guards an earlier carry-forward + 3s sentinel-wait regression). A naive "always re-extract for IDLE" fix would re-break it.
• Safer direction: persist the positive prepare-hook file association, and log a warning whenever a commit carries a trailer with no materialized checkpoint.
• Separately worth investigating: why Pi emitted no TurnStart for that rapidly-submitted third turn.

Verification

• mise run fmt / mise run lint → clean (0 issues).
• New test fails only at the invariant; premise and the two related existing IDLE tests pass.

🤖 Generated with Claude Code

---

Note

Low Risk
Test-only addition with no production or hook behavior changes; CI failure is intentional until a separate fix.

Overview
Adds an intentionally failing regression test in phase_postcommit_dangling_trailer_test.go that locks in the invariant: if PrepareCommitMsg adds an Entire-Checkpoint trailer, PostCommit must materialize that checkpoint.

The test simulates an IDLE session after condensation (FilesTouched cleared, prior LastCheckpointID set) where new edits exist only in the live Gemini transcript (no TurnStart, no shadow branch). PrepareCommitMsg still adds the trailer via live-transcript overlap with staged files; PostCommit currently skips condensation, so entire/checkpoints/v1 is never created and LastCheckpointID stays on the old ID—a dangling trailer. Comments point at a follow-up fix (e.g. persist prepare-hook association) without naively re-breaking TestPostCommit_IdleSession_NoTranscriptFallbackForCarryForward.

Test-only change; CI is expected to stay red until the behavior fix lands.

^{Reviewed by Cursor Bugbot for commit c511f36. Configure here.}

[Copilot]

Pull request overview

Adds a new strategy-layer regression test that reproduces a prepare-commit-msg vs post-commit mismatch where a commit can gain an Entire-Checkpoint trailer that points to a checkpoint that is never materialized locally (dangling trailer), particularly for IDLE sessions with empty FilesTouched.

Changes:

• Introduces a new (currently failing) regression test that sets up an IDLE session with no shadow branch and relies on live-transcript file extraction to trigger trailer insertion.
• Asserts the invariant that if PrepareCommitMsg writes an Entire-Checkpoint trailer, PostCommit must materialize the corresponding checkpoint and metadata branch.