fix(waterdata): get_ratings — don't abort on one bad feature, don't silently truncate

[thodson-usgs]

Problem

Two reliability issues in get_ratings:

1. One bad feature aborts the whole batch. The per-feature download loop catches (httpx.HTTPError, ValueError, OSError), but _download_and_parse → _raise_for_non_200 raises the module's typed errors — RateLimited (429), ServiceUnavailable (5xx), RuntimeError (other 4xx) — which are all RuntimeError subclasses, not httpx.HTTPError. (A feature missing its assets.data.href also raises LookupError.) So a single rate-limited/failed/malformed feature in a multi-site request escapes the loop and kills the whole call instead of being logged and skipped as intended.

2. Large result sets are silently truncated. _search sent limit verbatim and returned only the first page — no pagination, no clamp — despite the docstring describing limit as a "page size … (capped at 10000)." A query matching more than one page lost the remainder with no indication.

Fix

1. Broaden the per-feature except to also catch RuntimeError (covers the typed errors) and LookupError.
2. Clamp the page size to 10,000 and follow the STAC next link until exhausted.

Behavior note for (2): only the case it fixes changes — queries matching more than one page (or callers who set a small limit) now return all matches instead of the first page. Common queries (default limit=10000, fewer matches) return the same as before.

Verification (live API + mock)

• Pagination (live): _search(bbox=[-95, 40, -92, 42], limit=2) now returns 177 features (previously 2).
• Resilience (mock): with _download_and_parse patched to raise RateLimited, get_ratings(...) logs Failed to download / parse … and returns {} — skipped, not raised (pre-fix it raised).
• Happy path (live): get_ratings("USGS-01104475", download_and_parse=False) → 1 feature.
• ruff clean.

🤖 Generated with Claude Code