docs: run-record schema v1 additive extension for the ctx-bench runner

[saldestechnology]

Summary

Extends the benchmark run-record schema (v1) with the optional fields the ctx-bench pilot A/B runner needs to capture cost, turns, acceptance results, and diff-size normalization. All new fields are optional at the top level, so schema_version stays 1 and the $id is unchanged, per the documented additive versioning policy. Because the top level is additionalProperties: false, each new field is explicitly declared.

New fields

• agent (object): agent-side accounting from the harness (claude -p result payload) — session_id, total_cost_usd, total_tokens (required within the object), and num_turns (optional even within agent, since claude -p does not guarantee it).
• acceptance (object): the ctx-independent functional endpoint of a run — command, exit_code, passed, duration_seconds (all required within it).
• normalization (object): lines_added / lines_removed / lines_changed, the denominator for per-line score deltas.
• study_id, task_seed, generator_version, scorer_ctx_version (the offline scorer build — existing ctx_version remains the harness build), retry_attempt.
• metrics.gate_blocks and metrics.gate_block_recovered: gate-block counts and recovery signal.

run-record.md is updated in lockstep: field table, extended example record (keeps the drift test meaningful), and a versioning note that these arrived additively for the ctx-bench pilot runner.

Verification

• cargo test --all-features --test study_scripts — all 4 tests pass, including run_record_schema_and_example_stay_in_sync.

🤖 Generated with Claude Code