Add the development pilot record, pilot-readiness engineering and evaluator repairs

[ammarphp]

What this adds

• Development pilot record (docs/development/evaluation-study/pilot-record.md). 96 assignments: 12 task-bank tasks × 2 seeds × 4 arms, on one pinned host and one model.
  • Every assignment was sealed, scored and verified with no stop, about 18.4 USD of CLI-computed subscription quota against a 192 USD admission threshold.
  • A separate adjudication found that the subjects mostly behaved correctly. All 8 refusals on the out-of-range task were valid, and no swapped-legend case was a real role error. Meanwhile the mechanical evaluator left about two thirds of unsupported_claim outcomes unresolved, and most of its invalid verdicts were wrong.
• Pilot-readiness engineering:
  • a pilot approval kind that binds the schedule seed and broker limits;
  • the design budget;
  • a cost check for usage that reports no inference geography;
  • a census of lost launches before preflight;
  • per-process attribution of sandbox-denial reports.
• Evaluator repairs. Two rounds, each with held-out cases committed before the change, and read-only re-judgments of the sealed smoke and pilot into separate directories. Neither round met its acceptance bar; the record states this.

What it is not

Development evidence only. It gives no treatment-effect estimate, the per-cell samples are tiny, and the scoring rules remain provisional pending review.

Checks

On macOS, the staged export's full suite collected 7206 tests: 7139 passed, 53 skipped (optional dependencies and dev-only state), 14 xfailed, 0 failed. A Linux-like governance run, with Seatbelt reported unavailable, has 0 failures.

The exporter's evidence, agent-surface and publication checks passed, and the task bank builds from the staged tree. main will be fast-forwarded only after this PR's Linux test suite job is green.