Finish observing the draft PR checks and remove the r98 checkout

Snapshot prepared from evidence through 2026-10-19 23:16 UTC. The request is to determine whether the BuildTestEmbedded change caused test failures, then push the review fix only if the comparison and final native-test gates clear. The evidence does not demonstrate a PR-caused failure. The source is now pushed; GitHub checks are still running, so the work is not finished.

The live handoff is hf-2026-10-19-verify-and-push-the-two-buildtestembedded-review-fixes-after-provisioning-recovers. Its RUNNING report was acknowledged at 23:01:32 UTC, recorded in out/handoff-report-2302.log. A prior report attempt at 22:19 ended with RELAY_FORWARD_FAILED; that is historical and does not describe the current report state.

The evidence archive stores the contents of the local out directory directly under handoff-artifacts/r98. The full evidence archive and earlier handoff remain at the r98 evidence archive, the preceding handoff body, and the complete review patch. The earlier r62 evidence report, Issue 1560, and the PR are retained.

Comparison results

All 48 previously unpaired cases now have native first-attempt evidence. The six earlier controls passed on both sides. The ten alternating vacuum-maintenance runs produced 4 PASS / 6 FAIL on frozen main A (daf1d2ad5148ea8a6843a96f53298f2a133345b8) and 8 PASS / 2 FAIL on frozen candidate B (26382d5d preserved commit object, with the review patch); the tested binaries matched. The main-side reproduction is tracked in Issue 1560.

Across the original 41 comparisons, A had 39 PASS / 2 FAIL and B had 38 PASS / 3 FAIL. The checkpoint-maintenance and startup-recovery tests failed on both sides. One initial result differed: testOrdinaryShutdownRetiresHeldEventConsumerWithinExistingBound passed on A and failed on B while waiting for callback entry at line 196, before shutdown began. Ten later primary A→B pairs all passed on both sides; the separate context pair also passed on both sides. The original callback failure remains unexplained and is not being relabeled as a main-branch flake. This does not prove the PR can never cause a failure.

The parent decision is “no PR-caused failure demonstrated.” This clears the requested comparison gate; it does not claim every possible test was covered. The historical 259 unexecuted cases remain outside this comparison, and the full suite is not claimed green.

Source push and PR readback

The comparison worktrees were frozen at A daf1d2ad5148ea8a6843a96f53298f2a133345b8 and B 26382d5d preserved commit object. The reviewed source was rebased onto main 451ba11b10b5c86dab70d6425750e355e9abfc4f; final head 095a9282472c2f1b8b583410ad8d2c16d436eee7 contains the required coauthor trailer. The rebase range-diff says the first 26 commits are patch-equivalent and only the canonical terminal-status diagnostic differs in review commit 27.

After the comparison and final five-test gates passed, that head was pushed to source branch using an explicit lease against prior head 892e20d2039c3b582e519260a3e43d376f744dee. The PR readback confirms it is OPEN, draft, unmerged, at head 095a9282472c2f1b8b583410ad8d2c16d436eee7, and the description body exactly matches the saved text (SHA-256 6ff7a70a0b67de1bf4a58d052969d65adf3bb6d24c6fe86d73d166f5cf2a0a94).

The first gh pr edit attempt failed while reading the older repository.pullRequest.projectCards field, which GitHub rejected. A structured GitHub REST PATCH succeeded, and the exact body was read back. The workaround and readback are recorded in out/source-push-result.json and out/source-push-readback-verification.json.

Current CI and remaining work

The independent audit verified all five exact-head native first attempts as PASS: two direct results and the final three deferred-follower results from one native command. The audit is in out/final5-independent-audit.md and .json. Target 1 overlapped recorded foreign r88 control42 processes; this is retained as context, and no machine-isolation claim is made.

As of the PR readback at 23:15:37 UTC, all 16 normal GitHub test-shard checks were in progress. The required bld-build check had not yet been reported. A single read-only build-watchman 0.0.21 observer is running in tracked session 28417, PID 1403073 / start ticks 256816243. Its recorded mode is plain PR observation; it did not request a CI rerun or use a merge mode. The watcher log says the sampled base check suites were queued with zero runs and it would not rerequest them. CI remains in progress; there is no green verdict yet. No CI rerun is authorized.

Remaining work is to observe the existing checks under the standing policy, preserve terminal evidence, confirm the watcher exits after its terminal outcome and its recorded process identities are absent, and then remove this task’s clone and only .aibuildcaches entries whose names contain its r98 path. Cleanup has not been done.

Keep one local build at a time and require at least 10 GiB free before a build. Stop only processes recorded by this task. Do not raise timeouts, reduce iterations, weaken or skip tests, blindly rerun CI, merge, enqueue, publish, or deploy. The TinySite outage explains why the provisioning fix was made; it does not authorize deployment. New commits must end with Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>.