[likun666661]

Summary

• add machine-backed heavy-task workspace observation before the self-check gate and after the bounded repair attempt
• surface observed final-workspace mismatches as gate diagnostics instead of relying on raw task-text path parsing
• keep engineering diagnostics advisory after one bounded repair attempt so official verifier/scorer authority remains final
• add generic regressions that observed extra final-workspace paths remain diagnostic and do not become local rejection

Verification

• npm --workspace @maka/headless run typecheck
• npm --workspace @maka/headless run build
• node --test packages/headless/dist/__tests__/heavy-task-self-check-gate.test.js
• node --test packages/headless/dist/__tests__/task-agent-controller.test.js
• npm --workspace @maka/headless test (97 suites, 686 tests, 685 pass, 1 skipped)
• git diff --check

Benchmark context

• Full terminal-bench-sample@2.0 rerun with DeepSeek V4 Flash on this branch completed 10 trials, 0 errors, mean 0.700 (7 pass / 3 fail).
• polyglot-c-py passed after the diagnostic prompted cleanup of /app/polyglot/cmain and scratch compilation.
• build-cython-ext exposed why observation diagnostics must not become a second local verifier: task-shaped heuristics should be shown to the model once and then left to official verification.