Add the offline evaluation-study harness (synthetic engineering evidence)

[ammarphp]

What this adds

• benchmarks/governance/, the offline evaluation-study harness. It contains:
  • evaluation contracts and campaign-manifest verification;
  • an independent, standard-library counting-likelihood oracle and a synthetic development task family;
  • deny-default Seatbelt subject isolation with a fail-closed process census;
  • the custody broker and delivery guard over RAVEL's stage workers;
  • fake, Claude Code and Codex host adapters (the real hosts are exercised only with hand-written fixtures);
  • treatment manifests with behavioural identity checks;
  • a mechanical evaluator, prespecified analysis and cost planning;
  • the campaign runner and CLI.
• tests/governance/, about 2,600 tests.
• docs/development/evaluation-study/, the study's working records: slice design, decisions, blockers, plan, the oracle review packet, the G1 record and the 2026-09-25 incident record.

What it is not

No model has been run through the harness yet. Every campaign so far is a synthetic engineering campaign that uses the fake host. Nothing here is an agent result, a treatment effect or a physics measurement.

Why a pull request

tests/governance has never run on Linux; only macOS runs and a macOS simulation exist. This PR lets the test suite job, the first real Linux run, finish before main moves. Per docs/development/distribution.md, main is then fast-forwarded to this commit rather than merged in the web interface.

Checks run on the source commit (553db5c) and on the staged export, on macOS:

• the exporter's own checks (evidence 17/17, agent surface, publication);
• the staged tree's full test suite, in chunks: 5029 collected, 0 failing once a git checkout and hooks are present, as in CI.