docs: run-record schema — duplication metrics and per-run budget ceiling

[saldestechnology]

What

Additive, optional extension of the benchmark run-record schema (schema_version stays 1, same $id, per the documented versioning policy):

• metrics.duplication (optional object): production near-duplicate accounting for the run's diff -- base_pairs, final_pairs, delta, new_pairs, removed_pairs, plus the threshold / min_tokens parameters used. A pair is excluded iff both endpoints' files are under tests/; delta = final - base is signed, so deduplication shows up as a negative value.
• max_budget_usd (optional top-level number >= 0): the effective per-run budget ceiling -- the per-task override when one is configured, otherwise the study default.

docs/benchmark/run-record.md is updated in lockstep: field-table rows, the example record (now carries a duplication object and max_budget_usd), and a versioning note for the Phase 3 additive extension.

Why

This is the recording end of the ctx-bench Phase 3 duplication endpoint. In the pilot, ctx score's new_duplication floors at zero: it can only count duplication added, never duplication removed, which made the refactor/dedup task (t2) structurally unmeasurable -- an agent that eliminated existing near-duplicates scored identically to one that did nothing. A signed duplication.delta makes dedup work visible and rewardable. Scoping pairs to production code (excluding test-internal pairs) keeps the metric from being dominated by idiomatic test-fixture boilerplate, and max_budget_usd records the budget ceiling each run actually ran under so cost/timeout analysis doesn't have to reconstruct it from study config.

The production-scoping rationale (and a request for first-class support in ctx duplicates / ctx score itself) is tracked in #26. This PR only extends the record schema; existing writers and readers are unaffected.

Verification

Drift guard is green:

cargo test --all-features --test study_scripts
test run_record_schema_and_example_stay_in_sync ... ok
test result: ok. 4 passed; 0 failed

🤖 Generated with Claude Code