filter_relevant: clear TypeError mapping JS camelCase options to Python (parity with JS #7)

[mikkel]

Summary

Mirrors 255BITS/patchling #7 (ad02203) in reverse: filter_relevant raises a single TypeError before any model call when JS camelCase options are passed. The error names every bad key along with its Python translation.

Semantics verified in both codebases

• retries → max_attempts: JS retries counts extra attempts (src/filter.js:1372 attempts = retries + 1). Python max_attempts counts total attempts (relevance.py:1237-1259). Message: retries=N (extra attempts) -> max_attempts=N+1 (total attempts).
• retryBaseMs → retry_backoff: both are the base delay for exponential backoff (*2**attempt), with no jitter. JS uses ms and caps the wait at 30s (filter.js:1406). Python uses seconds and has no cap (relevance.py:1258). Message: retryBaseMs=M (milliseconds, base delay) -> retry_backoff=M/1000 (seconds, base delay). The defaults differ: 200 ms vs 0.5 s.
• ignore → skip_patterns: both use the same 30 default globs, matched case-insensitively against the full path (** crosses /, * does not; neither regex nor substring). A JS array replaces the defaults, and false, null and [] classify every path. In Python, None applies the defaults, a sequence replaces them (it does not extend them), and []/() classify every path. Note that JS null corresponds to Python [], not None. The CLI --filter-skip-pattern flag appends instead, but this message covers only the library call.
• onDecision, reasoningEffort: filter_relevant has no such parameters, so the message says not supported in Python filter_relevant and names no keyword argument. (After the call, FilterRelevantResult.decisions returns the decisions in input order, but it is not a keyword alternative. Reasoning effort is fixed at "low".)

Tests

• tests/test_filter_options.py vendors the JS SNAKE_CASE_OPTIONS map. JS-only keys are listed explicitly as unsupported, and the remaining keys are asserted to be exact inverses. It also checks one error naming all bad keys, the error being raised before the model callable runs, and pins the skip_patterns behavior: None means defaults, []/() mean every path, a custom list replaces the defaults, and matching uses case-insensitive full-path globs.
• Offline pytest -q tests in a venv with pip install -e .: 195 passed, 2 skipped (live LLM opt-in), 0 failed. LIVE_LLM was unset, and no model was called.
• CI Core API Tests on 28d25a4: success.

No merge, no publish.