feat(gitchameleon): ready the funded evidence run

[tcballard]

Outcome

Makes the GitChameleon evidence run operationally ready without claiming that the funded run has happened.

• pins voyage-4-large as the concrete naive_rag baseline over the same per-example Markdown corpus as As Decided
• names the product-backed experimental arm asdecided throughout CLI arguments, records, schemas, fixtures, generated filenames, and documentation; the retired rac label is deliberately rejected
• pins dataset commit 799a6a33e572a07a8985914e7251f5dea54b0ac4 and upstream scorer commit 3a1b6045a6b2a276bd24d715589cb041f8eccb93
• fetches the immutable raw dataset JSONL at the selected commit, rather than a mutable viewer projection
• adds bounded provider retries, one-example shakedowns, explicit overwrite protection, and provenance-checked resume
• records hashes of the exact prompt and grounding used for every completion
• refuses incomplete or duplicate paired resolution records before statistics
• records the execution contract as accepted GCB-ADR-0003 and freezes the first-run inputs in run-config.json

Evidence

• 71 repository tests passed on the funded-run implementation
• 20 focused GitChameleon tests pass after the arm rename
• Ruff check passed for all changed Python files (EXE001 ignored because repository Python entrypoints are intentionally mode 0644)
• all 3 GitChameleon decisions validate under As Decided v0.26.2
• live dataset fetch reproduced 328 raw rows at the pinned revision
• generated 328 per-example corpora / 1,968 decision artifacts; 1,968 valid, 0 invalid, OKF conformant
• generated 656 no_grounding + asdecided bundles; As Decided ranked the governing version pin first in 328/328 product-arm cases
• mechanical contribution preflight passed with no blockers

Explicit boundary

No benchmark result is claimed here. ANTHROPIC_API_KEY and VOYAGE_API_KEY are not present on this machine, and the locally built upstream Docker scorer still requires a functioning Docker daemon. The paid 984-answer run and upstream executable scoring remain the completion gate for #10.

Progresses #10.

Authored under my direction with Codex.