Promote C13 qualification to baseline 2.3

[KeyffMS]

Summary

Resolves the qualification-assumption work from #17 and prepares C13 for the next full C01-C16 run.

Baseline 2.3 decision

• keeps the documented C13 semantic assertion unchanged: SubagentStart injects additionalContext and continue=false is not relied on as a startup blocker;
• separates that semantic assertion from the two runtime limitations observed in C13 diagnostic run Complete distribution and release hardening #8: ephemeral parent-thread registration and project-scoped synthetic-agent spawning;
• permits a non-ephemeral retry only after the recognized collab spawn failed: no thread with id error before SubagentStart.

Final C13 transport

• first attempt remains a real aligned project-scoped fixture_agent under codex exec --ephemeral;
• fallback uses a separate disposable repository with the real project-scoped SubagentStart hook/config but no project .codex/agents child definition;
• the synthetic child is materialized only at disposable CODEX_HOME/agents/fixture_agent.toml;
• filename, declared agent name, requested agent type and hook matcher are all aligned to fixture_agent;
• the child must echo an outer-generated opaque proof that is absent from the root prompt;
• repository immutability, session cleanup, auth metadata, read-only sandbox, approval=never and model-tool network isolation remain fail-closed.

Full qualification

• mode=full now runs harness v6 with permission for the narrowly gated C13 fallback;
• mode=c13 remains available as a short probe but is no longer required before full qualification;
• C01-C12 and C14-C16 retain their existing qualification implementations.

Evidence materialization

• prepare_capabilities.py applies a deterministic baseline-2.3 C13 overlay after unpacking the stable template archive;
• the overlay updates C13 README/fixture/config/prompt/run-command metadata and recomputes hashes.json;
• a functional test materializes the complete archive into a temporary root and runs validate_capabilities.py against the result.

Safety / scope

• no production PlanAnvil core change;
• no sandbox/approval bypass;
• no privileged container/capability changes;
• baseline 2.3 does not mark C13 REPRODUCED before live evidence;
• relates to Analyze C13 SubagentStart assumptions under codex exec --ephemeral #17.

After merge and green post-merge CI, the next live run should be PlanAnvil Codex qualification from main with mode=full.