GEAK fallback can promote candidates rejected by native accuracy validation #1433

Hyperloom's GEAK fallback can promote a throughput result after native validation rejects accuracy. This reproduces on the former PR #1428 head ecf07f0b7bd8ffb2ae65bcafea1927e0275d1d8f using the actual Coordinator, native sweep and promotion methods with CPU fixture subprocesses.

With baseline accuracy 0.94 and current-best throughput 110, missing evaluation or fresh accuracy 0.2 can still lead to throughput 150 being promoted and a 50% validated gain recorded. Native quality controls reject both. A healthy-quality control and a failed-process control establish the intended positive and negative paths.

The route has two defects:

• Explore can emit per_variant_outcomes with REVERT / accuracy_drop or REVERT / accuracy_unavailable, no winner, and no output throughput while the task itself succeeds because it produced measurements. GEAK writeback treats missing throughput as inconclusive and invokes another harness.
• The fallback sweep drops fresh evaluation artifacts, and _validate_geak_via_geak_harness checks throughput and GEAK's claimed speedup without applying the native accuracy requirement.

The correction is published in existing PR #1428 at f89eedaeff3a4499d7f068394c05aff633bba841 and awaits review/merge. It:

1. Preserves the matching current-round native accuracy REVERT, closes the candidate with the existing no-promote tombstone, and withdraws provisional journey KEEPs.
2. Reads evaluation only from the fresh replay workspace through the existing native parse_eval_results function.
3. Requires valid fresh accuracy and the unchanged native drop tolerance when a positive baseline accuracy is recorded. Old result or throughput-summary scalar accuracy cannot substitute for evaluation artifacts.
4. Carries the accepted fresh accuracy into the optimization stack, which CLOSE already reads.

The same sweep has an independent return-configuration defect: it ignores authoritative accepted_config.env_map. The correction reuses the existing shared normalizer, with the actual shell transport correction published in AMD-AGI/GEAK#452 at 812819d82f6baf6ff9fc694eeaff33aba06a1ca5. Twelve real Hyperloom → copied GEAK adapters → Python-S decoder → child-environment controls pass. Parser-only round trips are insufficient.

Final immutable candidate validation: 20,719 full-suite tests passed, 43 skipped, eight expected failures and 15 passing subtests, plus whole-repository Ruff, formatting, REUSE and whitespace checks. Independent source review and 128 focused tests pass. The actual-source reproduction retains 13 corrected cases, fresh accuracy-retention controls, empty-map no-material closure and artifact-only fallback controls. Original vulnerable-source controls and all outputs are retained. No accuracy or performance threshold changed. Final receipt SHA256 688bc268081e0acd1c22725ef5d11fd99e8b9208e3e0f96d0940062226df1bec.

This establishes a current control-flow defect. It does not establish that this exact route caused the historical MiniMax KEEP with degraded accuracy, whose exact executable/state remain incomplete. It is also not a GPU performance or dashboard-recovery claim.

Evidence under /home/sapmajum/work/geak_ci_delivery_20260907/composed_validation/hyperloom_fallback_accuracy_audit:

• Original replay receipt SHA256 beed72623418c17dd16a4e5bf1f94ab26f744348d73fdede0d5083b962b1ca93.
• Corrected-source 13-case receipt SHA256 50412445696274f15429782d6c38f0d7c8a305ccb189036ba46eab1577d53173.
• Independent review SHA256 8d6579c855d937a436e99b1c58ff631d75d21d2449249a7ff6029ded0c74f778.
• Fresh-accuracy stack-retention receipt SHA256 1c10c10972f2e948c5dee9c737c902688222989de810fb6b68526f2508659aad.