fix(schema): do not invent max attempts

[reacher-z]

What / source

Changes the optional generation_args.max_attempts schema default from 1 to null and regenerates eval_types.py. Sources that do not report an attempt limit no longer materialize a claim that one attempt was allowed. Explicitly observed Inspect values remain unchanged.

Fixes #236.

Review lane

• Fast
• Needs a human — this is a material output semantics change across converters.

Design agreed in: #236, including collaborator confirmation that this is worth doing.

Checklist

• Validation is clean — 2 lm-eval aggregate records and 1 Inspect aggregate record, with 0 warnings
• Offline regression tests added; full test suite green: 759 passed, 20 skipped
• Ruff check clean
• Content spot-checked: absent lm-eval value is omitted; explicit Inspect value remains 1

The adapter report, registry, and source-row checklist items are not applicable because this changes schema default serialization and drops no source records.

Decisions & coverage

• Decision / where: schema_version
  Chose / instead of: keep the current version because the field remains optional and accepts the same integer values; only its non-required default annotation changes
  Confidence: high
  General: no
• Decision / where: generated model
  Chose / instead of: regenerate eval_types.py from the schema and run post_codegen, rather than hand-editing generated behavior
  Confidence: high
  General: yes
• Decision / where: explicit values
  Chose / instead of: preserve source-provided Inspect max_attempts values while omitting unobserved values
  Confidence: high
  General: yes

Coverage: 2 lm-eval source tasks to 2 valid records, 0 dropped; 1 Inspect source run to 1 valid record, 0 dropped. The lm-eval outputs contain no max_attempts key; the Inspect fixture that explicitly reports 1 still emits 1.

Commands run:

• uv run ruff check every_eval_ever tests
• env -u NO_COLOR TERM=xterm-256color uv run pytest tests -q
• uv run --extra inspect pytest tests/test_inspect_adapter.py -q
• real lm-eval and Inspect CLI conversions followed by validation at final collection/developer/model paths

AI assistance

OpenAI Codex assisted with implementation, tests, and validation. I reviewed the diff and the generated output.

[borgr]

CI was never run here either — the workflow was queued as action_required for a first-time contributor. Approved, and all four rows are green.

This closes #236. The change is right: a record carrying max_attempts: 1 asserted an attempt count no source stated, and the two new tests pin both directions — Inspect keeps a real 1 it read from the log, lm-eval omits the field entirely rather than inventing it.

Flagging the lane rather than the content: this edits eval.schema.json and the generated types, and it changes what existing converters emit, so by CONTRIBUTING's review policy it wants a human read and not only an agent pass — the diff is four files but the blast radius is every record with a generation_args. Labelled auto-review so the pass happens too.