[irvineoy]

This change protects evaluator state across the remaining FlyDSL task runtimes and makes task maintenance fail closed. It also repairs the 39 packages that failed or warned in the initial qualification of the retained tasks. Original scored cases, seeds, numerical gates, warmups and sample counts are preserved.

• Follow task-local candidate helper imports and reject known evaluator-state access and new packages/extensions that shadow protected Python modules. Freeze candidate sources and captured action evidence during semantic review.
• Add a repair-only quality-loop mode, selected-warning repair and bounded attempts. Repairs retain the original task/case/timing contract, preserve failed proposals and need a fresh validator PASS plus an independent review resolving every retained finding. They do not promote an optimizer result or report a speedup.
• Validate measured outputs, every contributing sample in the repaired Event tasks, changed-input replays and poisoned fresh single-call graphs. Keep task-owned checks self-contained and use the canonical benchmark helper for the shared capture interface.
• Fix operator/reference defects and missing public dtype, stride, routing, scale-buffer and output contracts. Sage uses residual FP4 components; the FP8 MoE uses an explicit protected staged baseline with the independent PyTorch oracle unchanged. These repairs change implementation and benchmark identity. Remeasure both arms; historical scores are not comparable, and this PR claims no optimization gain.

Validation:

• CPU: python3 -m pytest -q tests — 19,196 passed, 195 skipped, 35 subtests passed. After the final unscored MoE boundary additions, focused checks passed again: 97 direct controls and 52 task-contract/history regressions. make check-perf-helpers also passed. The Linux process-cleanup test now handles a process disappearing while its /proc status is being read; its full test module passed (45 tests), followed by 20 repeated runs of both real subprocess cleanup cases (40 checks). This is a test-only race fix; task and evaluator runtime sources are unchanged.
• GPU: All 39 repaired task packages have framework-finalized PASS reports. Qualifications use MI355X/gfx950 and a pinned Docker runtime. Finalized reports, source/runtime identities, raw action results and earlier WARN/FAIL evidence are retained outside the task packages. Reports span immutable qualification snapshots; this is not a claim that every task ran under one Git snapshot.
• Sage: the formally qualified final implementation also passed 70 input checks at the unchanged 0.25 normalized-max-error gate; the worst observed value was 0.0642947. Extra FP4 compute/storage is part of the repaired baseline, not a same-cost accuracy or speedup claim.

No agent optimization campaign was run for these repairs. GPU qualification on other architectures is not claimed.

The source checks are structural controls, not a Python security sandbox. Submitted candidates still need their own full correctness/performance evaluation and independent semantic review.