Cross-tool efficiency: per-corpus reference artifact + docs

[Mathews-Tom]

Summary

Generates and checks in the per-corpus cross-tool efficiency reference artifact and documents the number. Builds on #366 (the tokens-at-fixed-recall metric).

What this adds

• benchmarks/cross-tool-efficiency/cross-tool-comparison.json: the reference artifact, produced from a clean run of archex benchmark cross-tool over the benchmark task set. Metric values are not hand-edited.
• docs/LOCAL_METRICS.md: a "Cross-tool efficiency (offline benchmark)" section citing the artifact and its per-corpus figures, stating the number is offline benchmark-only and never enters the in-process ledger or archex metrics summary.
• A test asserting the checked-in artifact parses, grades the localization family as its own corpus, holds recall equal in every scored comparison, and renders.

Per-corpus reference values (tokens at 100% required-file recall)

Corpus	Naive model	Comparable	archex	naive	Reduction
self	full_file	16/24	9,484	4,416,681	99.8%
self	grep_window	16/24	9,484	2,626,845	99.6%
external-comprehension	full_file	16/19	22,681	783,725	97.1%
external-comprehension	grep_window	16/19	22,681	492,119	95.4%
external-localization	full_file	20/21	13,247	469,836	97.2%
external-localization	grep_window	20/21	13,247	408,410	96.8%

"Comparable" counts only tasks where both paths reach 100% required-file recall, so no figure compares unequal recall. Localization is graded as its own corpus, never merged with comprehension.

Artifact location

.archex/ is gitignored repo-wide (and by a user-global ignore), so the committed reference artifact lives under the tracked benchmarks/ tree rather than punching negation holes in the ignored workspace dir. The archex benchmark cross-tool default output stays .archex/cross-tool-efficiency for ephemeral ad-hoc runs.

Out of scope (unchanged)

• No in-process metrics ledger, metrics reporter, retrieval ranking, or default change.

Stack

Stack-Id: cross-tool-efficiency-cfdfb5
Base: feat/cross-tool-token-model
Position: 2/2

1. feat/cross-tool-token-model -> Cross-tool token-efficiency: naive baseline + tokens-at-fixed-recall metric #366
2. feat/cross-tool-artifact-docs -> this PR

Depends on: #366

Validation

• uv run ruff check / ruff format --check / uv run pyright on changed Python — pass
• uv run pytest tests/benchmark/test_cross_tool.py tests/benchmark/test_reporter.py --no-cov — 80 passed
• Doc figures verified byte-for-byte against the checked-in artifact aggregates