fix: align task contracts with observable verification

[zfj1998]

Summary

This draft PR proposes task-contract and verifier-quality repairs found while
studying whether skills learned on T1--T3 transfer to changed T4--T6 inputs.

The core concern is attribution: when a semantically correct implementation is
rejected by an unpublished label, exact phrase, arbitrary tie-break, or
source-code spelling check, apparent "skill learning" can actually be verifier
adaptation.

The full rationale and per-task findings are in
docs/task_contract_quality_audit.md.

Scope

Across the 90 held-out T4--T6 tasks, this patch:

• repairs 27 outcome or visible task-contract issues;
• rewrites process-only checks for 27 additional tasks so they exercise
  observable behavior instead of private source spelling/layout;
• leaves 36 held-out tasks unchanged;
• replaces two scheduling instances whose original scoring depended on hidden
  arbitrary choices (one had no positive-length feasible intersection); and
• retains all 30 skill-family capability targets.

The patch includes only benchmark task assets and focused quality regressions.
It intentionally excludes model harnesses, experiment-platform integration,
watchers, and analysis/reporting infrastructure.

Main defect classes

1. Instruction/verifier mismatch — output paths, allowed null markers,
   reply obligations, assignee forms, or result cardinality differed.
2. Hidden arbitrary contracts — exact labels, raw list order, slot IDs,
   fixed phrases, and one-reference-token checks were not published.
3. Impossible fixture — an E6 scheduling task required a common window that
   did not exist.
4. Source-shaped process checks — tests searched for variable names,
   helper-file locations, regex tokens, or comments instead of running the
   promised behavior.
5. Cross-task policy drift — six inbox-triage tasks used inconsistent
   priority/deadline boundaries and occasionally omitted the decisive blocker
   from visible fixtures.

Validation

• python scripts/validate_assets.py: PASS, 30 families / 180 tasks
  (the local review environment emitted expected warnings because Docker/Harbor
  were unavailable).
• pytest tests -q: 15 passed.
• Modified JSON files parse successfully.
• Modified Python files compile successfully.
• Ruff and git diff --check pass.
• Checked-in reference solutions were also executed in real task containers:
  90/90 held-out T4--T6 tasks passed strict verification.
• A fresh live-agent family smoke on the final fixtures passed all six T1--T6
  tasks.

The reference audit establishes internal solvability, not that every task is
perfect or that a particular model should pass. Live runs remain useful because
they can expose valid output surfaces that the reference solution does not
exercise.

Review requested

This is opened as a draft because the change is intentionally audit-sized.
Feedback would be especially useful on:

1. whether the published policies match the intended latent skills;
2. whether process verifiers should constrain implementation structure at all;
3. whether semantic equivalence or a published deterministic tie-break is
   preferable where several answers are valid;
4. whether the replacement scheduling instances preserve the intended
   difficulty; and
5. whether maintainers would prefer the patch split by environment or defect
   class before merge.