Build: measure focused-engine source-to-result latency

[chrisbbreuer]

Parent: #494
Depends on: #462
Related: #617

Goal

Measure and gate the smaller production focused-engine artifact against the combined inline unit-test artifact with comparable source-to-result cache boundaries.

Scope

• Version the build-feedback artifact rather than rewriting the accepted V1 history.
• Add isolated cache groups for library, focused-engine, combined-unit, and TSan scenarios.
• Measure focused frontend cold link/run and cached runtime selection beside equivalent combined-unit cold/cached paths.
• Measure focused-engine and combined-unit TSan paths separately.
• Preserve the full cold/warm unit-suite phases and exact shard plan.
• Validate exact focused denominators and nonzero combined-unit denominators in every retained sample.
• Remove completed cache groups during collection so the evidence job stays within bounded disk usage.

No-workaround rules

• No warm-cache result labeled cold, shared cache pollution between compared artifacts, omitted full-suite phase, reduced sanitizer behavior, zero-match pass, outlier removal, dirty-worktree publication, or evidence collected under overlapping build/test work.

Acceptance

• V1 evidence remains accepted and reproducible; new collections use an explicit V2 schema.
• Comparable focused and combined cold/cached/TSan boundaries are machine-readable and fail closed on denominator or process-quality drift.
• Three sequential samples on a clean exact revision retain every raw command, output, resource measurement, and full-suite plan.
• Focused source-to-result latency improves materially while the full combined integration gate remains exact.
• The Build: reduce compile, link, and test feedback time #494 documentation and checked evidence identify the precise scope; 0 Test262 behavior changes.

[chrisbbreuer]

Implementation is on main:

• 7eb51f12 adds the V2 ten-scenario matrix, isolated cache groups, exact denominator/result validation, V1 migration coverage, and bounded completed-group cleanup.
• c3f186f2 corrects the singular one-case output contract and preserves incomplete raw artifacts when final validation fails.

Validation is green: direct/build-integrated collector self-tests, dependency audit (15 classified edges, 0 migrations), 267-file docs link audit, and checked 49-page docs build.

No performance evidence is promoted yet. The first one-sample diagnostic correctly failed final validation on the singular grammar mismatch. The corrected repeat established valid cold denominators but was deliberately aborted when an unrelated user-owned Bun/Home Zig test graph began compiling concurrently; that host load invalidates timing evidence. The reproducible temporary cache was removed, and no samples were discarded or retried into a claim.

#614 remains open for one clean diagnostic followed by the required three-sample publication run when the host is idle. Flips 0 Test262 cases.

[chrisbbreuer]

Fail-closed overlap handling is now complete on main:

• fa89331c rejects another Zig compiler/build driver before each scenario and preserves all orchestration failures as incomplete raw artifacts.
• 483d0fa1 closes the start-after-preflight race by repeating the same process-boundary audit after every measured command.

The live integration fixture ran while the user-owned Home build was active. V2 rejected before clean_library and wrote exactly:

• schema: zig-js-build-feedback-v2
• complete: false
• requested samples: 1
• collected rows: 0

The fixture was removed after inspection. This is validation of evidence rejection, not timing evidence. The clean one-sample diagnostic and three-sample publication remain pending until the external suite stops and the host settles.

[chrisbbreuer]

Evidence-safety hardening pushed in d8fc8c97 (fix(bench): reject contaminated build feedback (#614)).

The V2 collector now shares #617’s exact process classifier, rejects unrelated Zig/maker, Home test, unit/corpus, and frontend benchmark processes both before and after every scenario, records complete-process CPU occupancy, and fails below a 60% minimum. V1 validation remains unchanged for the checked historical artifact. The fail-closed policy lives in one shared module so exact-parent and focused-build evidence cannot drift independently.

Validation: build-feedback self-test passed; exact-parent self-test passed; dependency audit remains 15 classified edges / 0 migration targets / 60 tools. This changes tooling only and flips 0 Test262 cases. The clean three-sample V2 collection remains pending an exclusive host window.