Phase 2: base squash/diff on the captured post-rebase bump tip (T9)

[petemoore]

Phase 2 of the rework (docs/WORKPLAN.md): the T9 narrow fix. Confirmed empirically in Phase 0 (see docs/WORKPLAN.md → Validation experiments).

The bug (T9)

squashBranch reset the agent's work onto the scan-time pr.HeadSHA. The Phase-0 rebase rewrites the branch head, so that SHA is stale; git reset --soft <stale> then produces a "fix" commit whose diff bundles every unrelated main change merged since the branch's original base. On petemoore/taskcluster #193 the "fix" commit is 300 files / +27,957/−42,840 for a ~3-file query-string compat fix.

The fix

Keep inherit-and-rebase (Q11 → A); only change the base SHA the squash uses.

• Capture the real tip in cloneAndBranch right after checkout -b branch pr.HeadRef, before the worker runs — the true post-rebase bump commit — onto a new Pipeline.bumpTipSHA field. rev-parse failure is now fatal (this SHA is load-bearing).
• Single canonical base (M4): bumpTipSHA is now the base for the squash and the reviewer diff/commit listing. The reviewer previously diffed origin/<HeadRef>...HEAD — mis-based, and reads empty/wrong once the worker force-pushes (MINOR-1). Both now use the stable captured commit.
• Plumb it to the orchestrator (MAJOR-1 / N4): added RunResult.TipSHA, set on every exit path via a named return + one defer. The gave_up path records its terminal outcome — and the sticky comment's SHA marker — against TipSHA (via terminalSHA(tip, scan)), so the next scan's SHA-skip fires on the rebased head instead of re-entering the expensive agent. Every other recordOutcome call stays on pr.HeadSHA per the agreed scope.

⚠️ Phase 0 finding (flagged, not blocking)

#193's bloat is dominated by the -X theirs rebase reverting 289 unrelated files into the bump commit — a broader form of the Q11 residual. This narrow fix is correct and necessary but does not by itself clean a #193-style branch (it bases off that same polluted bump tip). Recorded in docs/questions.md for a maintainer decision; out of scope here (would relitigate Q11).

Tests

• TestSquashBranchUsesCapturedTipNotStaleBase — git integration: builds bump + 2 agent commits, squashes on the captured tip, asserts the fix commit is parented on the bump and contains only the agent's files; demonstrates the stale base would bundle the bump file.
• TestTerminalSHA, TestGaveUpSkipFiresAfterRebase — the N4 SHA selection and that the skip fires on the rebased head.

Local: go build/vet/test ./..., gofmt, staticcheck all clean.

🤖 Generated with Claude Code