fix(core): main-red — align audit FAIL/WARN cache-put pins with cycle-1594's declared-content contract - #519
Merged
Conversation
…ycle-1594's declared-content contract Main's go workflow has been red since 20e839e (cycle-1594's autonomous ship). That ship deliberately changed worktreeContentSHA from `git add -A` to `git add -u`: untracked-unstaged files are residue, never adopted into the audit-bound tree or the verdict-cache key (5 pins in phase_bindings_staging_scope_test.go + eval gitignore-staging-sweep). The change is judged CORRECT and is kept — it closes the "binding certified content no auditor reviewed" class from cycles 1572/1574. What broke: the lane's gate does not run `-tags integration`, so it missed phase_bindings_fail_verdict_test.go, whose two pins seeded worktree dirtiness as an UNTRACKED changed.txt. Under the new contract that delta is residue → the tree keeps base identity → the ADR-0048 fresh-base guard refuses the verdict-cache Put: - TestEmitPhaseBindings_AuditWARN_CachePut_Control FAILed on both platforms — the vacuity control doing exactly its documented job ("if this control ever reds, the FAIL pin's no-cache assertion is vacuous"). - The FAIL sibling's no-cache-put assertion had indeed gone vacuous: the fresh-base guard, not the verdict guard, was doing the skipping. Fix: both tests now modify the TRACKED committed f.txt (an unstaged tracked modification — explicitly pinned as captured content by TestWorktreeContentSHA_CapturesUnstagedTrackedModification), restoring the verdict guard as the ONLY discriminator between the FAIL and WARN pins. Production code untouched. Swept initBindingRepo consumers: no other test still relies on untracked-file adoption. Triage of the OTHER main-red sightings, so nobody chases them again: - "--- FAIL: TestPersonaStopCriterionDedupe_CombinedLineCountReduced" in CI logs has NO duration — it is stderr echo of stubSelfCheckRunner fixture text from build_persona_budget_check_test.go, not a real failure. - aadab95's red was TestRunPool_BackfillsReplacementWhileSiblingLaneStillRunning (internal/fleet, 0.00s) — not reproducible in 5 local -race runs and green on the next main run; classified flake (first sighting of this class). Follow-up class (queued, not absorbed): lane test gate misses integration-tagged tests, so a lane can ship a semantics change that reds main CI only. This is the second inherited-red burn (#518 inherited this one). Bar by EXIT CODE: gofmt clean, vet exit=0, go test -count=1 ./... exit=0, -race -tags integration ./... exit=0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BgEQLTgfon8RSVN9bQHBtU --- ## Actual diff (v8.34.0+) Files modified (1): - M go/internal/core/phase_bindings_fail_verdict_test.go 1 file changed, 7 insertions(+), 2 deletions(-) Reviewed-by: code-simplifier Reviewed-by: go-reviewer
mickeyyaya
deleted the
fix/main-red-cacheput-control-declared-content
branch
September 1, 2026 07:41
mickeyyaya
added a commit
that referenced
this pull request
Sep 1, 2026
…hip rejoin the lane gate; the retake is the contention answer (#523) Root cause of the cycle-1594 red-main class (inbox lane-gate-misses-integration-tags, 0.9): internal/core, cmd/evolve and internal/phases/ship were put on integrationTierEnvExclusive on 2026-07-19 (8e2afef — contention false-REDs, cycles 930/931/932), and the serialized retake-on-red that properly cures that exact contention landed ONE DAY LATER (3c5ed71, 2026-07-20). The list was never revisited, and a green test (TestEnvExclusive_ExistingEntriesUnchanged, "must stay env-exclusive") pinned the fossil — a test that copies the config it guards detects an edit, never an error. Cost realized on cycle-1594: a lane changed internal/core, its stale integration-tagged pins ran NOWHERE lane-side, CI caught it only post-merge, and main stayed red 2.5 days (20e839e; #519 fixed the tests; #518 inherited the red). The list shrinks to internal/bridge alone — the only entry with a REAL residual reason: its requireTmux tests boot tmux sessions that time out under a live wave AND skip in CI (no tmux on runners, #483), so neither the retake nor CI can vouch for it; its honest backstop stays a quiet-host run. For the three returning packages the flow is: red -> serialized retake under the cross-lane exclusive lock -> red-then-green absorbed as a visible WARN, red-twice a real offender FAIL. TDD red-first: TestIntegrationTierScope_CoversCoreCmdShip_Cycle1594 FAILed with the fossil's own skip message before the change. Fossil pins inverted (TestEnvExclusive_OnlyBridge); the skip-semantics test repointed to bridge; the backstop-note helper's CI-affirmative branch stays pinned via a hypothetical entry. code-simplifier caught and fixed a stale preceding comment paragraph that still asserted the old membership — the same duplicated-belief class, in prose. Architecture review BLOCKed the first cut of this very fix — for spreading the disease it cures: the replacement belief was prose-restated in six places and three copies were already stale inside the diff. Remedy applied: envExclusiveEntry{pkg, why, backstop} is the single record; the emitted WARN, the lane-log note, and the whole-suite filter all PROJECT from it; the quiet-host-only map and suffix helper dissolved; the selection criterion ("excluded only when the retake cannot vouch AND CI has no backstop") is pinned as a RULE over the table (TestEnvExclusive_EntriesDeclareNoCIBackstop, anti-vacuity floor included) with the Cycle1594 test kept as the labeled instance regression; a serialized-retake deadline kill now degrades to the fail-open WARN instead of masquerading as red-twice offenders (quiet-host tier measured ~127s vs a fixed 15m budget and a 3.6x-7.7x contention multiplier); and the false cmd/evolve premise (fleet-soak is in-process fakes, not real tmux) is corrected in the record's doc. The re-review BLOCKed the remedy itself — the retake's governing prose still forbade the deadline-WARN the new code performs (the same duplicated-belief class, introduced by the prior CRITICAL's own fix) — and found the deadline path discarding flushed offenders plus untestable behind a const budget. All applied: the governing trade-off block now splits exec-failure (attempt-1 offenders, never laundered) from deadline-kill (evidence outranks the budget — offenderMarkerLine/hasOffenderMarker is the one home of the marker belief, fallback-free, so flushed verdicts in a truncated retake still FAIL and only a marker-free truncation degrades to WARN); integrationTierTimeout became a var (apicoverTimeout's stated rationale) with a deadline orchestration test PAIR beside the existing three; the WARN now carries each entry's why; the "NOT covered by CI" criterion phrase is a hoisted const referenced by data and rule test alike; ghost literals are annotated as reintroduction guards. go-reviewer separately APPROVEd, independently proving the CI-backstop premise from go.yml and re-timing the three suites (119/81/62s loaded) against the per-attempt budgets. Bar by EXIT CODE: gofmt clean, vet exit=0, internal/phases/audit full suite -count=1 ok. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BgEQLTgfon8RSVN9bQHBtU --- ## Actual diff (v8.34.0+) Files modified (5): - M go/internal/phases/audit/ciparity.go - M go/internal/phases/audit/ciparity_integration_test.go - M go/internal/phases/audit/ciparity_unit_test.go - M go/internal/phases/audit/envexclusive_bridge_test.go - M go/internal/phases/audit/integration_tier_orchestration_test.go 5 files changed, 283 insertions(+), 157 deletions(-) Reviewed-by: code-simplifier Reviewed-by: go-reviewer
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Main's go workflow has been red since 20e839e (cycle-1594's autonomous ship). That ship deliberately changed
worktreeContentSHAfromgit add -Atogit add -u(declared-content contract: untracked-unstaged files are residue, never adopted into the audit-bound tree). Two integration-tagged pins inphase_bindings_fail_verdict_test.gostill seeded their delta as an untracked file — residue keeps base identity, so the ADR-0048 fresh-base guard refused the cache Put and the WARN control red both platforms. The lane gate never runs-tags integration, so the lane shipped it green.The production change is correct and stays (5 pins + eval
gitignore-staging-sweep.md); the stale tests are the defect. Both now modify the trackedf.txtcommitted byinitBindingRepo, with comments pinning why the delta must be a tracked modification.Verification
acs/cycle1495, safe — pre-stages via its ownadd -A)go test -count=1 ./...exit=0;go test -race -count=1 -tags integration ./...exit=0 (full module, both tiers)gofmtclean,go vet ./...exit=0TestRunPool_BackfillsReplacementWhileSiblingLaneStillRunningred classified as flake (5/5 local-racePASS); the deterministic red entered at 20e839eFollow-up class (queued, not absorbed)
lane-gate-misses-integration-tagsinbox item filed (P1, 0.9): the lane gate must cover-tags integrationfor touched packages — second inherited-red burn (#518 inherited this pair).🤖 Generated with Claude Code
https://claude.ai/code/session_01BgEQLTgfon8RSVN9bQHBtU