Timing supplement (13 Nov 2026)

This supplement is separately versioned from the frozen v0.9.2 correctness results. Its observations are environment-scoped candidates. It does not add a certified latency corpus to the snapshot.

Cold case observations

The retained reports contain 1170 rows: 1085 timed and 85 untimed. Outcome counts are 338 inconclusive, 393 not-reached, 354 reached, and 85 unsupported. Timing absence for unsupported cases is not zero. Identity differences against the frozen reports: 0; outcome changes: 0. These checks do not add analyzer support or alter correctness scores.

Each row below summarizes one cold invocation per case. The distributions describe cases within each slice; they are not repeated measurements of one case and do not estimate runtime variance. The denominator counts timing sidecars regardless of outcome. Values are complete adapter-declared invocation wall-clock milliseconds; setup, builds, fixture materialization, normalization, validation, and hashing are outside the timer. Quantiles use linear interpolation over the per-case observations. No overhead is subtracted.

On narrow screens, scroll the table horizontally to see every column.

Cold case distributions, milliseconds
Slice	Rows	Timed n	Min	Median	p90	Max
Bifrost C kernel	58	58	559	598	631.1	1157
Bifrost C++ kernel	68	68	591	608.5	634.3	1102
Bifrost C# kernel	70	70	593	613	635.1	1000
Bifrost Go kernel	70	70	697	724.5	758.4	1178
Bifrost Java kernel	70	70	1349	1380.5	1405.5	8356
Bifrost Java modeling	24	8	1379	1396.5	3449.5	8136
Bifrost Java native	12	0	—	—	—	—
Bifrost JavaScript kernel	70	70	596	610	631	980
Bifrost JavaScript modeling	24	8	591	612	733.6	987
Bifrost JavaScript native	12	0	—	—	—	—
Bifrost Kotlin kernel	70	70	1386	1462.5	1568.5	8156
Bifrost PHP kernel	70	70	595	611	634.2	1010
Bifrost Python kernel	70	70	591	605	625.2	967
Bifrost Python modeling	24	8	593	608	737.4	995
Bifrost Python native	12	0	—	—	—	—
Bifrost Ruby kernel	70	70	588	602.5	621.1	976
Bifrost Rust kernel	64	64	655	679	722.6	1089
Bifrost smoke (mixed languages)	118	117	603	629	1405	7748
Bifrost TypeScript kernel	70	70	598	613	630.1	1009
Infer C kernel	56	56	254	263	281	964
Infer C++ kernel	68	68	258	268	1627.2	1821

Invocation overhead

These are fixed cold invocations of the existing trivial no-flow fixture. The full min-to-max range is an upper-bound estimate that includes startup and the fixture's own analysis. It is not warm marginal cost and is not subtracted from the cold case values. The groups remain unqualified.

Separate invocation-overhead estimates, milliseconds
Adapter	Three repeats	Min–max range
Bifrost / Python no-flow	805, 565, 562	562–805
Infer / C no-flow	245, 260, 260	245–260

Qualification and provenance

The continuous queue passed its initial quiet-host admission. After dispatch, CPU contention and unavailable observations were retained as telemetry while the queue continued; this does not certify uncontended timing. The original A24 inter-adapter settling qualification was not established, and descendant containment remains unproven. All cold slices and both overhead entries are marked uncertified. 63 of the planned cold groups remain blocked. Only Bifrost at v0.15.1 and Infer at v1.3.0 have new cold observations; warm marginal measurements and the other analyzers remain blocked.

The mistaken Bifrost PHP overhead arm was rejected before policy analysis or a timed repeat. A separately source-reviewed continuation corrected it to the existing A24 Python definition and ran the still-unstarted Infer C arm. It repeated no cold groups. The rebuilt runner has its own pinned digest and is not byte-identical to the unavailable earlier binary. Tool/runtime identities, source projection, fixtures, configurations, resource limits, and environment qualifications are described in the immutable reviewed campaign record .

Evidence and downloads

There are no accuracy changes and no cross-release comparisons. The candidate observations do not support a speedup claim. The full evidence bundle is separately addressable from the frozen release:

Save the three lossless bundle files together before verification or restore. The compressed index is served with an opaque suffix to preserve its bytes; save it as files.jsonl.gz. The bundle's relative paths begin with snapshot/. The reviewed campaign record includes the verification and restore commands.

Frozen release commit: 86092336d8f9ca32fe9e2c4042c2dced76c1feb1.
Published report manifest SHA-256: 2d4a40a10c79813a4dad8e2b8f7340daa6259da0553762a92c5847ec5d91014c.