fix(headless): require validated verifier grades

[me2seeks]

Summary

• make rejected verifier payloads authority-neutral instead of treating field presence as a grade
• validate every verifier attempt index so sparse arrays fail closed
• cover malformed, sparse, and internally inconsistent runner output while preserving provider infrastructure classification

Follow-up to #1591.

Verification

• npm run build:test
• npm run lint
• npm run format:check
• npm --workspace @maka/headless run typecheck
• npm --workspace @maka/headless run test:dist — 1,429 passed, 4 environment skips, 0 failed
• git diff --check

Review focus

This changes only verifier-derived scoring authority. Existing completed-cell, benchmark-deadline, and tool-step eligibility behavior remains unchanged; a verifier influences classification only when structuredVerifierGrade() accepts the full outcome/reward/attempt contract.

[Astro-Han]

[P2] Reproject stale fallback scores when the stored verifier is rejected

readFixedPromptWal() sends persisted task_completed events through projectStructuredVerifierOutcome(). If structuredVerifierGrade() rejects the stored verifier, the event is returned with its old score flags unchanged.

I reproduced this with a legacy failed max_tokens event written by the previous verifier-presence fallback. scored: true and eligible: true survive resume, the runner is not called, and the event still counts toward the acceptance denominator. The malformed-WAL test currently starts with both flags set to false, so it misses this case.

Please recalculate the identifiable failed/no-deadline/invalid-grade legacy records with the same rules used for fresh output. The existing authority for tool_step_cap_reached, completed cells, and deadlines should stay intact.

Approving with this P2. I would hold the merge until it is addressed.

[Astro-Han]

Btw the E2E test has been fixed on main. Rebase to main would solve the Red CI.

[me2seeks]

Addressed the replay P2 and rebased this branch onto current main.

Fresh output and stored-WAL replay now use the same scoring projection. A structured verifier contributes scoring authority only after structuredVerifierGrade() validates it. For rejected legacy verifier data, replay only repairs failed, no-deadline rows that are identifiable as output from the old verifier-presence fallback. This keeps completed cells and benchmark deadlines authoritative, while tool_step_cap_reached remains eligible but unscored without a valid grade.

The regression coverage now includes missing, non-array, sparse, and reward-disagreeing attempts; the stale passed: true shape written by the old fallback; policy_denied; and preservation of tool-step-cap, completed, deadline, and no-verifier records.

Validation on the updated head:

• npm --workspace @maka/headless test — 1,304 passed, 4 skipped
• node --test packages/headless/dist/__tests__/fixed-prompt-controller.test.js — 89 passed
• npm --workspace @maka/headless run typecheck
• Biome check and git diff --check

The earlier E2E conflict is also gone after the rebase.

[Astro-Han]

Review: MERGE-READY ✅ (2 independent reviewer passes)

DeepSeek review of the full diff including the incremental commits after the earlier approval (head 08354425):

• Prior P2 (reproject rejected stored verifier grades) is properly addressed: the read-time reprojection uses the same taskCompletedScoringProjection as fresh writes, and idempotency is verified (728-combination corpus: 0 non-idempotent).
• Legacy-reprojection discriminator is precise: exhaustive 728-combination audit (status × 13 errorClasses × reward × verifier shapes × deadline) shows exactly the 27 fallback-produced rows change (scored→unscored), zero collateral damage to completed/deadline/valid-grade/no-verifier rows; every presence-fallback row is caught.
• Sparse-array validation closed: the new index-explicit loop rejects holes anywhere (including trailing holes via at(-1)), closing the old .some()/.slice().some() bypass. Zero effect on main-era stored data (JSON serializes holes to null) — purely defensive.
• Verified state: typecheck clean, full headless suite 1308 pass / 1 env-skip / 0 fail on merge-tree with current main, biome clean, git diff --check clean. CI green.

Non-blocking P3 (optional):

1. The discriminator keeps {scored:true, max_tokens, no verifier} rows as-is (differs from fresh path) — unreachable from main-era writers, conservative choice, worth a comment.
2. Fresh-path policy_denied/tool_step_cap_reached + invalid-verifier variants and setup_failed/verification_error replay remap lack direct tests (same code path, low value).
3. PR branch typechecks red against its own stale fork base (provider-request-trace.ts — fixed upstream by #1679, not this PR's fault); verify against current main instead.

None block merging.