ci: qualify explicit native Release producer profile

[davidgmbb]

Closes #604.

Summary

• pass -DBUSTER_DEBUG_INFO=OFF to both ordinary native generation and missing-cache recovery
• preserve the existing non---ci Release policy: BUSTER_CI=OFF, -O3 -DNDEBUG, frame pointers, tests, unity/single object, the default linker, and semantic flags
• add a network-free policy regression that rejects --ci, disabled frame pointers, or drift between the two generation paths
• document the exact hosted reproduction profile without changing local Release profiling defaults or the newer --jobs 4 differential policy
• retain the four-platform screening audit

Exact submitted revision

Head ddce1b86e6201e632e10a826f6a524462e2f6f3a is one focused commit on parent 6db3c81a331544a60b859fa5a4b6fcb19db8d642. The diff is exactly four files: workflow policy, regression test, reproduction guide, and screening audit.

Hosted pre-publication validation passed:

• python3 tools/native_producer_profile_test.py -v
• python3 tools/differential_ci_policy_test.py -v
• python3 tools/new_audit.py --check
• git diff --check
• exact changed-file and documentation-anchor checks

Qualification result

The issue-specific gates are complete on the exact submitted SHA.

• Guarded generated-compiler throughput: run 35392054038; 40/40 valid benchmark pairs, identical-output enforcement, zero confirmed regressions, gate exit 0. Nine timing cells remain explicitly inconclusive rather than being relabeled equivalent.
• Three full native attempts: 35392053996, 35392532813, 35392539715; all four Unix lanes passed mode matrix, full differential, result/reproduction, packing, and retained upload in every attempt.
• Three self-host/fixed-point attempts: 35392053972, 35392534747, 35392541872; all passed bootstrap, repeated generation/fixed point, compiler regressions, independent compiler-oracle checks, and evidence retention.
• Exact-head cold telemetry: run 35393099586; three isolated builds per platform, 12/12 successful, byte-stable binaries per platform, exact cache/argv verification, deleted-cache recovery parity, peak RSS, package, and upload timing retained.
• Native workflow comparison: run 35394188368 against exact-parent baseline 35391693255; all 16 compared native jobs passed. Candidate median native critical path was -16.08%; aggregate native runner work was +1.99%. Hosted variance is retained transparently, so this PR does not advertise a causal whole-workflow speedup.

The attributable same-cohort screening result remains a 14.2%–21.8% median producer-build reduction and lower peak build RSS on all four hosted Unix platforms while preserving the required compiler and diagnostic contract.

The detailed evidence, platform telemetry, timing tables, artifact references, and interpretation limits are recorded in the qualification comment on this PR.

Merge status

All exact-head PR workflows completed successfully: Buster CI, guarded compiler throughput, self-host fixed point, native-retirement contract, GPU toolchain acceptance, and iOS monitor lifecycle. The PR is open, non-draft, conflict-free, and mergeable. It has not been merged automatically; ordinary repository review policy still applies.

[davidgmbb]

#604 exact-head qualification complete

Qualification is complete for submitted head ddce1b86e6201e632e10a826f6a524462e2f6f3a (tree 60de82d429dc3ba84faedd380a015a10aeee0625) against exact parent 6db3c81a331544a60b859fa5a4b6fcb19db8d642.

Production policy

The patch changes only the native hosted producer policy:

• ordinary generation: add -- -DBUSTER_DEBUG_INFO=OFF
• missing-cache recovery: add the identical override
• retain BUSTER_CI=OFF, -O3 -DNDEBUG, -fno-omit-frame-pointer, tests, unity/single object, default linker, semantic flags, and local Release profiling defaults
• add a network-free policy regression that rejects --ci, disabled frame pointers, or drift between the two generation paths

Guarded generated-compiler throughput

Run 35392054038 passed on the exact base/head pair with guard enabled and identical-output enforcement:

• 40/40 benchmark pairs valid
• 0 confirmed regressions
• 9 statistically inconclusive timing cells; these are retained as inconclusive, not relabeled equivalent
• maximum reported median wall delta +0.47346%
• maximum reported RSS delta +0.21659%
• gate exit 0

Three exact-head full-native attempts

All four Unix native lanes succeeded in each attempt, including execution-mode matrix, complete differential, result/reproduction, packing, and retained upload:

1. 35392053996
2. 35392532813
3. 35392539715

That is 12/12 successful exact-head native jobs and four retained native artifacts per attempt.

All three matching self-host/fixed-point attempts also passed ordinary bootstrap, repeated generation/fixed point, compiler regressions, independent compiler-oracle checks, and evidence retention:

1. 35392053972
2. 35392534747
3. 35392541872

Exact-head cold-build/RSS/recovery telemetry

Run 35393099586 checked out the submitted SHA directly and completed three fresh isolated profile-C builds on each hosted Unix runner. All 12 builds passed exact cache/argv checks, retained one optimized unity compile plus one link, produced byte-stable binaries within each platform, passed mode execution, and preserved the same profile after deleting CMakeCache.txt.

Platform	Median complete build	Sampled peak RSS	Evidence upload
Linux x86-64	148.158 s	1,781,133,312 B	1 s
Linux AArch64	107.402 s	1,956,278,272 B	1 s
macOS x86-64	141.025 s	1,672,601,600 B	1 s
macOS AArch64	75.531 s	2,592,505,856 B	2 s

The RSS value is the retained child-process ru_maxrss sample; nested-process RSS is not deduplicated.

Whole native-job timing and runner work

Run 35394188368 compared the exact-parent debug-info-on baseline 35391693255 with the three exact-head attempts. All 16 compared native jobs were successful.

Metric	Baseline	Candidate attempts	Candidate median	Median delta
Native aggregate runner work	2064 s	2066 / 2105 / 2309 s	2105 s	+1.99%
Native observed critical path	939 s	780 / 788 / 966 s	788 s	-16.08%
Native pack work	50 s	47 / 57 / 57 s	57 s	+14.00%
Native upload work	10 s	9 / 7 / 8 s	8 s	-20.00%

Complete native-job medians versus the single parent baseline were Linux AArch64 -16.47%, Linux x86-64 +35.48%, macOS AArch64 +27.37%, and macOS x86-64 -15.72%. This large cross-run variance is why the qualification does not advertise a causal whole-workflow speedup or add independent lane savings together. The isolated same-cohort screening remains the attributable build-profile comparison: profile C reduced median producer build time by 14.2%–21.8% and peak build RSS on all four platforms.

Result

The three remaining gates recorded in the retained screening audit are satisfied on the exact submitted tree:

1. guarded compiler throughput with identical-output enforcement;
2. three complete exact-head native + self-host attempts on all four Unix lanes, with cold-build/RSS/recovery/evidence timing;
3. complete native critical-path and aggregate-runner-work comparison without double-counting savings.

Copilot review has also been requested through both the documented REST reviewer and gh pr edit --add-reviewer @copilot paths; it is not counted as qualification evidence unless GitHub actually submits a review.

Issue-specific qualification is complete. This PR remains unmerged pending the repository's ordinary final check/review policy.