Handoff: decide fixture step deadlines and finish the independent-resolver PR gates

Written 2026-10-04 10:53 UTC.

RE-VERIFY: This is a write-time snapshot. Read gh pr view https://github.com/CodexCoder21Organization/kompile-buildscript/pull/87 --json state,headRefOid,statusCheckRollup,mergedAt and git ls-remote origin 'refs/heads/wip/*87*' refs/heads/wip/L3-callback-audit refs/heads/fix/resolve-outside-compile-locks refs/heads/main before acting. The old 05:52 source/decision snapshot was superseded. The original full body is preserved in initial-live-handoff.md.

Mission

The owner requested: “shared caches must have fast, reliable locking internally so that serializing externally is never necessary; where that does not hold, we need concrete failure reproductions.” This handoff originally required choosing explicit callback lineage or rejecting pending custom-recipe admission because a joined callback waited on its own producer. Later revisions removed custom callback sharing and completed result reuse altogether, retaining main's resolver call sequence. The original ownership choice is now unnecessary; the published PR still needs its current source/fixture review and green required CI. The supervisor owns merge, enqueue, deployment, publishing, handoff completion, and the fixture scope decision below.

What was found and done

1. L3 claimed this exact handoff for fable-hq-20261004 and fetched saved/live bodies. Live PR https://github.com/CodexCoder21Organization/kompile-buildscript/pull/87 is OPEN at d67a59446ca1c24505d6ccefc4836a332dcc4686, based on main 5a81c5854192c9d2a2536662efbb86f9f3694f6c. The handoff's ffe128eb PR snapshot was stale. Its current description opens with the original owner's request and explicitly removes custom result/preparation reuse.
2. Original mechanism verified: outer resolver joins its prestarted executor's nested compile; the original nested compile joins the shared custom recipe future. ed175d6 removes that custom sharing/completed reuse and invokes CoordinateResolver directly before package, context and entry compile locks. The nested call resolves independently. Private builtins still share their own immutable results/flights. No implicit callback-lineage detection is claimed.
3. The retained callback test/helper/wrapper are byte-identical between 11646e4 and d67 (git diff is empty). Original B87b XML is 277/278, callback FAIL 11095 ms. Later B87e exact-head d67 XML is 284/284, callback PASS 4216 ms and exact resolver sequence PASS 4205 ms. L3 additionally rebased and ran both existing public tests locally without changes: 2/2 PASS, 0 FAIL, callback 11936 ms and sequence 13429 ms. Full errors, exact-head comparisons and separate test-comprehensiveness/adversarial reviews are in L3 findings. No L3 product change or PR edit was made.
4. L3's selected remote run e3d7ae01 progressed through TESTING and finished FAILED; client reports 1 failed test, returns no XML/details. WUI results fail view materialization with HTTP 503 and the 5-second cached-view error; raw-run watchman cannot find the run among 100 reported runs. The failure is retained explicitly and is not dismissed because local tests pass. No CI rerun was requested. Raw client log and API/watchman responses are durable.
5. Current published CI remains red: GitHub all-tests FAILURE; kotlin.build remote FAILURE; Gradle SUCCESS; optional kompile-remote-build IN_PROGRESS. Current callback failure dump shows nested Kotlin compilation, not the old custom recipe future wait. Artifact filename 0.0.44 alone does not prove which code bytes loaded; the newer repair adds candidate-byte verification.
6. An existing active B87f repair started 10:04 UTC, with an 11:54 UTC deadline, and owns the same PR's remaining source/CI work. Its latest remote checkpoint 117e12e contains deterministic stale-input/retained-map witnesses and passing-target logs, but product source in that branch is still d67, not the locally reported passing fix. Its notes report5/5fixed witness/trace/close and 12/12 fixture passes. Source issues: preparation hashes script B before locking, but compilation rereads script C afterward; same-thread callback returns a copied partial map instead of main's live read-only view. Do not infer those fixes are published from prose; demand its clean product checkpoint/patch and new exact-head suite. L3 did not edit this overlapping repair.
7. Scope finding requiring supervisor decision: B87f fixture diff replaces nested.get(10,TimeUnit.SECONDS) and outer compilation get(20,TimeUnit.SECONDS) with untimed get, removes the callback's stack-dump timeout catch, and similarly removes 5/10-second successful completion and 20-second release bounds in other helpers. New nested-resolver admission, nonnull compiled output, candidate-byte verification, and isolated fixture resources are useful additions, but do not preserve those removed constraints. This lane's brief says “Never increase a timeout ... weaken an assertion” and “If you believe a test is wrong, stop and write that up under NEEDS_USER with evidence.” An unchanged 25-second child / 30-second test limit does not preserve a removed inner 5/10/20-second assertion. The concrete proposed fixture diff is reviewable; L3 has not adopted it.

Latest B87f update (live ref verified again at 10:52 UTC): its corrected classpath witness now restores exact B JAR bytes, fails old source with fresh B returning C, and passes the fixed source 1/1. Notes report final four product controls 4/4 and two remote full suites in flight; the later suite covers the strengthened witness. The product commits named in those notes still have no remote product checkpoint. This is additional passing evidence, not a green required suite or a published fix. The notes explicitly say those remote command sessions have no outer deadline; the supervisor should ensure its current repair owner bounds and cleans up both waits. L3 owns neither session and stopped neither.

Relevant PRs and refs

Repository for every row is kompile-buildscript. No source, artifact, or production deployment was published by L3. No branches or PRs were closed. Keep the current PR and preserve archives until the supervisor finishes content/evidence review; older custom-sharing candidates must not overwrite the current resolver contract.

PR: https://github.com/CodexCoder21Organization/kompile-buildscript/pull/87, OPEN d67a59446ca1c24505d6ccefc4836a332dcc4686; required checks red. Keep and finish the current repair, not the obsolete 11646e4 candidate. Open unrelated PRs https://github.com/CodexCoder21Organization/kompile-buildscript/pull/85 and https://github.com/CodexCoder21Organization/kompile-buildscript/pull/90 do not replace this change; L3 made no recommendation to alter them.

Next steps

1. Reverify current PR head/checks and the active B87f remote refs first; another repair may have advanced since this snapshot. Read L3 findings and B87f's latest findings/review as evidence, distinguishing actual product bytes from notes.
2. Supervisor question: must all existing fixture step deadlines remain, or may the PR-added fixtures be deliberately redesigned to rely on the unchanged outer 25/30-second limits? Recommendation: keep stated bounds by default and investigate the measured failing path. If the operator deliberately permits the redesign, record that explicit exception, require complete output/callback-count assertions and candidate-byte provenance, and retain a stack-dump watchdog. The general testing documentation's new-test sizing guidance does not override this lane's unconditional prohibition. Neither choice accepts red CI or blames host load.
3. Let the existing PR repair owner finish and durably push the independently proven prepared-input and live callback-map source fixes; review the exact clean tree. Salvage useful class-byte, nested-admission, output and per-test-resource improvements while following the supervisor's deadline ruling. Do not reintroduce custom recipe flights or result reuse.
4. After rebase on latest main and exact-head full remote suite, update the still-open PR, state why all fixture changes exist, and use tracked plain build-watchman to finish required kotlin.build (remote) and GitHub all-tests. Do not re-run a failed check for luck. Capture full stack/results for the unexplained remote e3d7ae01 failure before using it as diagnosis.
5. Supervisor performs final review and decides merge/handoff completion only after the applicable gates are met. No merge, enqueue, deploy, Maven publication, handoff complete or release was authorized for this lane.

Reusable / operational knowledge

• Fresh clone: /code/ws/L3/kompile-buildscript. Evidence-only worktree: /code/ws/L3/kompile-buildscript-evidence, branch wip/L3-callback-audit. No unique product work exists only locally; unchanged verification source is remote PR d67.
• Export PATH=/code/ws/bin:$PATH in every shell. Repo's current scripts use a checksum-pinned JVM Coursier launcher themselves; no shim was needed. rg is absent on this host; grep was used after checking it.
• Exact selectors: scripts/test.bash --remote --test kompile.buildscript.testResolverCallbackReentersSharedRecipeOnSeparateThread --test kompile.buildscript.testCallerResolverSequenceMatchesMain --log <path>. L3's local fallback ran only those two existing tests after cgroup memory dropped, with a 20-minute outer process cap. Never run a full local suite on this shared 6 GiB host.
• CLI's run summary and WUI may disagree about available result records. Raw failed summary is not a passing test gate, a no-details remote failure is not a root-cause diagnosis, and a pending/not-found watcher result is not proof an executor died.
• Watchman's base ruleset gates only bld-all-tests here; the lane separately treats kotlin.build (remote) as required. Optional kompile-remote-build being pending does not change a red required gate. Never pass --to-merged.
• The original B87b analysis, invariants, before/after proofs, fallback/generation diagnostics, exact classpath ordering, transport failure/recovery, two-process compilation and borrowed-output close evidence remain remotely preserved; see the original body and archived branch table. No evidence is discarded by this shorter rewrite.
• All L3 test/watchman/report command sessions ended; final process/worktree audit is recorded in findings. No other lane's process was stopped. The claim remains with fable-hq-20261004 because the brief forbids release except ALREADY_CLAIMED or CHECKPOINT; completion belongs to the supervisor.