PR15040 maintainer takeover: head-specific nonvisual proof
Repository: pingdotgg/t3code
Original head: 6a40eb390c434de9741506a71ce8f8f161de4c90
Review base: 54983e3f458871dffe755ea908ffe0d60cea101b
Integrated top: c0e1ff56523da29dc780db9ebe90a0f8555d3e87

Layer heads:
scheduler-tasks-05/plan-reads c00b86c85e38eb44bcc00c55df93ef8e8be0296e
scheduler-tasks-05/claude-homes 7b797e30c16fb7768344225822a8d3e5d073a2bf
scheduler-tasks-05/acp-resume cc2e2ad065c6f13125080ac0ee5522cdc5628523
scheduler-tasks-05/watch-recovery f5fe7661e47fbc888c045dc541fc5ede20358331
scheduler-tasks-05/server-services 9b701f0f35df689045e00b9e75fc2c1839a24ab0
scheduler-tasks-05/first-turn-branch 0d95f074aab8c7e641d6deedc9083db61e9710be
scheduler-tasks-05/title-refinement c0e1ff56523da29dc780db9ebe90a0f8555d3e87

Observed before / after:
Initial minimal tests produced 12 expected failures and 2 new passing controls against the old pinned base.
R1 regressions: symlink repoint after target resolution read new settings once but left the new target unwatched; real Git first-turn naming renamed an inactive temporary ref and changed metadata while HEAD stayed user/selected.
Both added R1 cases failed on unchanged production code; 9 controls passed.
Fixed: recovery attaches and checks the whole topology; later new-target edits reach cache/subscribers. Native status guard preserves user-selected HEAD, original temporary ref SHA, and thread metadata.
R3 before/after: malformed unrelated Claude instance blocked healthy script reads before; independent decode/skip now preserves healthy custom, environment, and default roots and escape checks.
One extra receipt-replay parameter proves finite completion and rollback when a successful dispatch makes no projection progress.

Final integrated verification:
core: Test Files  4 passed (4); Tests  77 passed (77)
watch: Test Files  2 passed (2); Tests  93 passed (93)
services: Test Files  2 passed (2); Tests  32 passed (32)
branch: Test Files  5 passed (5); Tests  206 passed (206)
acp: Test Files  1 passed (1); Tests  8 passed | 113 skipped (121)
claude-runner: Test Files  1 passed (1); Tests  4 passed (4)
claude-native: Test Files  1 passed (1); Tests  1 passed | 297 skipped (298)
Current workflow/transcript rerun: 7 passed, including healthy content with invalid unrelated config.
Current total: 415 unchanged integrated results reused +7 fresh workflow/transcript passes =422 distinct passing cases. Repeated executions are not added. Unselected cases are not failures.
Server, contracts, and provider-acp TS7 checks: exit 0, zero errors; advisory suggestions remain.
Lint on the 28 changed files: exit 0, 14 existing warnings. git diff --check: exit 0.

Proof method and limits:
Real temporary Git repositories and production driver; persisted SQL/orchestration/effect-worker milestones with fake provider delivery.
Three independent RpcTest clients exercise the exported production shared handler layer and one maintenance process; no real socket/auth upgrade graph. Route scope and shared scanner/registry lifetime are source-verified.
Incoming main SDK query-runner tests use real installed SDK/synthetic history through a Linux worker; its already-landed implementation is not part of this replacement diff.
No UI change. No browser, native Windows/macOS/iOS/Android, packaged Electron, older Node, real account-based turn, relay, or tunnel verification.
Automatic first test-launch refresh added three cached packages with zero downloads; tracked manifests/lockfile stayed unchanged.

Scope comparison: 6 original claims already fixed on pinned main; 8 remaining behaviors in 7 layers.
Replacement adds 27 cases in 10 declarations across 8 existing owner test files; no new test files.
Independent review: R1 findings fixed; R2 superseded by main integration; R3 config-isolation finding fixed after 1 red /1 valid control; R4 Opus and Sol both clear on every layer and the cumulative diff; both independently reran workflow/transcript 7/7. Three completed parallel rounds (R1, R3, R4); R2 cancelled/superseded before verdict.
Local verification only; current-head CI is reported separately after publication.