mirror of https://github.com/rcourtman/Pulse.git synced 2026-10-26 13:39:47 +00:00

Files

T

pulse-triage[bot] 92c3297049 ci: add fixed UUID layout diagnostic without release gate waiver

The exact release tree retains a failed paired UUID benchmark check, while push CI cannot repeat that comparison. Provide a reviewed-source route to collect the requested four-condition hosted layout evidence instead of repeating local samples or changing product order.

Fix source/tree/toolchain identities, isolate diagnostic controls, alternate ten rounds, and retain partial receipts without publishing binaries. Execution still needs separate operator authority; collection success does not dispose of the failed gate. Nine focused harness tests and existing benchmark contract tests pass.

Change-source: pulse-maintainer

2026-09-06 22:11:03 +01:00

5.0 KiB

Raw Blame History

Fixed UUID layout diagnostic — not release qualification

This main-only workflow collects evidence for the failed PR1943 UUID comparison in run 34055039318. It does not change the benchmark gate, its threshold, product declaration order, release checks, or the candidate. Green here means collection completed, not that the candidate passed its failed check.

Exact experiment

Baseline 5845692daf302a1524c2ca8fe6d06339ae7f8a24, tree dedef52efe4fc28d1c05c9d957a0cc439f54db22.
Candidate 57f3b6401a6553efaa7646d5f1db446a2725b9dc, tree 7c2765cb06f424ce49647b33a5ebf3ade375e828, identical to the failed synthetic PR merge ad6869a29cd0b54c918967d3cf41320914ba5790.
Ubuntu 24.04 hosted runner, Go 1.26.7 linux/amd64, no toolchain auto-upgrade or restored build cache, GOMAXPROCS/test.cpu=4.
Original baseline/candidate and the same two trees with normalizeSegment moved immediately before normalizeRoute, without changing its body. Both HTTP source and benchmark files must match their recorded SHA256 first.
Four full API test-package compilations before measurement, identical frontend embed stubs. No unit tests run. Only ^BenchmarkNormalizeSegment$/^uuid$ executes, including warm-up.
Ten rounds, forward/reverse condition order alternating, 300ms per sample, one invocation per condition per round. No retries, early statistical stopping, broad benchmark selection, or automatic verdict. A failed invocation stops collection while preserving partial evidence.

This reproduces the design of Core's retained 6 September controlled experiment, not its local measurements. Hosted CPU and layout effects are still hypotheses. Contemporaneous paired/interleaved samples reduce time-drift bias; they do not eliminate shared-host interference. See Go performance monitoring guidance.

Execution authority: unresolved dependency

Source integration alone does not authorise Delivery to dispatch this workflow. Delivery's standing dispatch permission covers only Pro license/relay; starting the release assessment service is not permission to run arbitrary Actions. Do not use either mechanism to smuggle this experiment into execution.

After ordinary source review and main integration, an operator with repository Actions dispatch authority must explicitly approve and perform one execution of uuid-layout-diagnostic.yml on main, supplying its full reviewed main commit as expected_workflow_sha. Alternatively the operator must explicitly grant a bounded execution route. No such grant is supplied by this document. The workflow rejects branch/SHA mismatch and Actions reruns (run_attempt != 1); a new manual dispatch is not automatically a justified retry. Record the approval and dispatch run identity internally. Before any further attempt, reconcile the first run and its partial artifacts and obtain a reasoned diagnostic decision, not a green-run search. A new source pair or toolchain requires source review, not free-form inputs.

Evidence and disposition

Download the one uuid-layout-diagnostic-<run>-<attempt> artifact. Require metadata.json to say complete: true, verify SHA256SUMS, and verify all four conditions have ten UUID-only samples. Retain raw per-invocation samples, aggregates, chronological load/order records, exact source/tree identities, workflow SHA/run/attempt, Go environment, CPU/kernel observations, binary hashes, HTTP-only symbols/disassembly, and the four HTTP source files showing the intervention. Experimental executables and source archives are temporary and are never uploaded, packaged, tagged, installed, or proposed as candidates. Partial artifacts cannot establish a completed experiment.

Reuse the original hosted artifact (archive SHA256 61acc9f4e14fc0672f1eacaa81481e2ec497c71d302fe2bca30bc598c2a5354d) and Core's already-retained controlled results; do not overwrite or substitute them. Compare original and reordered pairs separately with a recorded benchstat version, retaining all results including adverse ones. The historical result is +21.04%, ten samples; a smaller later number alone is not causal disposition.

The release qualification owner must then explain whether the intervention establishes a defensible disposition of the unchanged candidate or requires a reviewed repair. Existing adverse evidence and the HOLD remain until that judgment. This diagnostic neither investigates nor clears the separately excluded SQLite work. No release assessment is warranted merely because this workflow exists or completes.

Focused harness checks

python3 scripts/tests/test_uuid_layout_diagnostic.py uses fake subprocesses to check fixed scope, ordering, source/toolchain rejection, partial evidence, non-overwrite behaviour and receipt completeness without executing benchmarks. For any real local build/benchmark execution use pulse-heavy-run -- <command>; a further local sample is not the missing hosted evidence.

5.0 KiB Raw Blame History