Clean-baseline patch replay as the official grading path + fresh-execution release gates

[ElbertePlinio]

Make clean-baseline patch replay the official grading path — the agent's workspace is never what gets graded. Panel-unanimous prerequisite for any published multi-model leaderboard.

• Replay the captured patch (app/lib/core/patch_capture.dart) into a fresh baseline copy before hidden grading; stop grading inside the agent workspace
• Grade-time assertion that hidden fixtures were never readable from the agent workspace (isolation asserted in the release gate, not only at workspace assembly)
• Release gates verify fresh execution: current hidden-test digests, admission cleanliness, sandbox-enforced provenance (release_report.dart today mostly re-checks stored claims)
• Fold the manual tasks/flutter/*/deepswe/ replay artifacts into runner behavior, or archive them

Note: patches + manual replay artifacts already exist, so this is a moderate lift, not a rebuild.

[ElbertePlinio]

Parent plan: #10

[ElbertePlinio]

Grading-critical note from diff_size review (2026-07-09): patch capture (app/lib/core/patch_capture.dart) runs 'git add -N .' + 'git diff --binary' — unstaged only, so fully staged agent changes are MISSING from the captured patch. Harmless for diff telemetry, but under clean-replay grading the captured patch IS the graded artifact — staged-only work would silently vanish. Fix in this issue's scope: capture 'git diff HEAD --binary' (staged+unstaged union) and add a staged-only regression test.

[ElbertePlinio]

Second grading-critical note (from #4 review, 2026-07-09): runtimeIsolation.generatedCodeSandboxEnforced is set from 'sandbox object non-null' (task_qa_runner.dart ~:634), i.e. config-on, not proof of at least one successful sandboxed process. When building fresh-execution gates here, assert enforcement from an actual successful sandboxed prepare/run, not object presence.

[ElbertePlinio]

Follow-up pinned from PR #29 (Codex P2, comment 3557670302): the release gate must validate corpus-manifest entries, not just presence. Specifically, before a report is 'ready': each benchmark.selectedTasks entry's taskId/taskVersion/taskBundleDigest must match the corresponding leaderboard.tasks row, and corpusManifestDigestSha256 must equal a digest recomputed over those entries (not just be 64-hex). This is part of #6's fresh-execution verification.

[ElbertePlinio]

PR #31 opened (clean-replay grading + fresh-execution gates). Deeper hardening deferred to #30.

[ElbertePlinio]

Shipped in PR #31 (squash-merged 4f7cd27, CI green). Clean-baseline patch replay is now the official grading path — the agent's workspace is never graded. Hardened across 5 Codex review rounds; deeper follow-ups tracked in #30.