Add repository repair replay and verify native index readiness

[acgxv]

The native benchmark searched after only eight worker passes, which left larger source packets partly indexed. Drain the dedicated benchmark queues before retrieval, and return an explicit error for failed jobs or a bounded drain timeout.

Add a repository repair pilot with three frozen historical defects, two repetitions, and four conditions: no memory, file search, model-written Git notes, and native ELF recall. Every condition uses the same six-turn agent and isolated behavioral checks. Invalid JSON consumes a turn and receives feedback; final tests determine success.

Preserve both the initial 24 diagnostic trials and the full corrected 24-trial replay, with source/image fingerprints, contexts, patches, per-request costs, and failures. Corrected outcomes are 5/6 without memory and 6/6 for the other conditions. This small calibration does not establish an ELF advantage over file search. Cumulative API cost, including the earlier screen, is $0.094396915 against the $10 ceiling. Update the runbook, evidence index, and OpenWiki.

Validation:

• 60 benchmark contract tests and repository tooling tests passed.
• Rust/TOML format and documentation checks passed.
• Each old source failed the evaluator; each historical reference repair passed.
• Rebuilt native Rust adapter; live readback verified 222/222 index jobs DONE with zero failures. Cold and warm retrieval each returned context for all six cases.
• All 48 trial receipts and provider usage are retained. Run-owned containers, volumes, and networks were removed.