Preregister agent qualification studies and v2 execution controls

[Taz33m]

What changed

• publish the completed v1 agent-qualification protocol, prompts, runners, and tests
• preregister the held-out C4/C5 generalization cohort without publishing evaluator-only identities or gold
• add a separately frozen execution layer with exact-order enforcement, provider-specific authorization, native skill isolation, technical verdicts, and fail-closed first-party Claude routing
• publish the completed v2 aggregate outcome with hashes binding it to the preserved private scorecard
• keep all cases, gold, transcripts, workspaces, and authorization records under the ignored private tree

Why

This makes the agent-qualification work reproducible without leaking evaluator-only materials or rewriting the preregistered protocol after measurement. The public result is deliberately narrow: it reports safe-pass counts, required breakdowns, claim boundaries, and the machine-enforced follow-up motivated by the failures.

Completed v2 outcome

All 24 preregistered runs received valid pre-score technical verdicts before semantic grading.

• public docs: 1/12 safe passes (8.3%)
• frozen skill: 7/12 safe passes (58.3%)
• skill minus docs: +50.0 percentage points
• skill-side unsupported candidate-defect accusations: 0/12
• compatible-engine case: docs 0/6, skill 1/6
• profile-boundary case: docs 1/6, skill 6/6

This exceeds the preregistered +20-point directional threshold, driven mainly by boundary handling. It is not presented as a statistically conclusive model comparison, production-adoption result, or proof of reliable autonomous compatible-engine qualification. The frozen cohort will not be revised or rerun.

The aggregate report is in docs/agent-qualification-generalization-v2-results.md. The machine-enforced evidence follow-up is intentionally separate in #71.

Verification

• 49 focused experiment tests pass locally
• Black passes on the experiment runners and focused tests
• Python 3.10-3.13 CI is green on the preregistration/execution commits
• the frozen v2 execution completed with 24/24 valid pre-score verdicts
• private scorecard and narrative hashes match the values recorded in the public results note

Privacy and scope

Evaluator-only case identities, gold, provider authorizations, transcripts, workspaces, per-run outputs, and grading remain ignored. This PR publishes protocols, runners, aggregate outcomes, and audit bindings only.