feat: diff_scope assertion — measure fix scope (files/lines changed)

[slowdini]

Motivation

A core claim of investigating-bugs is "make the smallest targeted fix; don't bundle
refactoring." That is a scope property, and llm_judge rubrics grade it unreliably. eval-magic
already materializes the fixture in env/, so it can diff the before/after and expose objective
scope metrics.

In the timezone eval, the correct source fix is small and single-locus; a per-consumer nudge
spree (or a symptom-patch-plus-cleanup) is larger and touches more files. A diff metric makes that
difference measurable instead of judged.

Proposed surface

A new assertion type plus benchmark reporting:

{
  "id": "minimal_fix",
  "type": "diff_scope",
  "max_files_touched": 1,
  "max_lines_changed": 8
}

And always expose the raw metrics per run in benchmark.json (even with no threshold), e.g.:

"diff_scope": { "files_touched": 1, "lines_added": 3, "lines_removed": 1, "hunks": 1 }

Optionally make the metrics available as context to an llm_judge rubric (so a judge can weigh
"is this the smallest correct fix" with the numbers in hand).

Semantics

• Diff is computed over the case's files set plus any files the agent created in env/.
• Thresholds are an opt-in assertion; the default behavior is to report the metric, not gate
  on it.

Acceptance criteria

• benchmark.json includes diff-scope metrics per condition.
• An optional diff_scope assertion can PASS/FAIL on max_files_touched / max_lines_changed.
• Documentation states plainly that "smallest" ≠ "best" — hard gates are a secondary signal,
  to be combined with a correctness assertion, never used alone.

Back-compat

Additive. No change to existing suites unless they add the assertion.