fix(connectors): honor --limit + fix gzip/taxon-id/multi-word terms (uniprot, pdb, clinvar)

[001TMF]

What

Three connectors (uniprot, pdb, clinvar) ignored Query.Limit and walked/fetched the full result set even for a count query (--limit 1) — causing timeouts — and each had a connector-specific bug that made benchmark/count queries fail outright. Found while wiring these sources into the multi-domain retrieval benchmark.

All three now mirror the ncbi-virus limit-as-walk-ceiling pattern (#55). Complete-or-fail is preserved: the strict limit < authority guard makes a limited walk always BestEffort with AuthoritativeCount = upstream total, so a limit can never produce a false Complete.

uniprot

• gzip: readBody returned raw gzip bytes (UniProt gzips its JSON; it arrived undecompressed), so json.Decode failed with invalid character '\x1f'. Now decompresses on Content-Encoding: gzip or the 0x1f 0x8b magic sniff (covers both search and getEntry).
• limit: honors Query.Limit as a walk ceiling; exposes the authoritative total on the BestEffort path.

pdb

• organism_taxon_id → HTTP 400: mapped to rcsb_entity_source_organism.ncbi_taxonomy_id, which is search-disabled upstream. Remapped to rcsb_entity_source_organism.taxonomy_lineage.id (string in), which the RCSB Search API accepts.
• limit: honors Query.Limit as a walk ceiling and caps getEntry to the limited id set (was fetching every entry of the full total even for --limit 1).

clinvar

• limit: honors Query.Limit as a walk ceiling on the esearch idlist.
• multi-word terms: quotes Entrez values containing whitespace (e.g. "Uncertain significance"). Unquoted multi-word values were mis-tokenized by esearch and silently returned 0.

Verification

• Per-connector regression tests added, each proven non-vacuous by revert.
• Live counts confirmed against each source's own number (e.g. ClinVar BRCA1 Pathogenic = 14,186 matches esearch exactly; UniProt human reviewed = 20,431; BRCA1 VUS 0 → 7,877 after the quoting fix).
• make ci green; govulncheck clean.
• Adversarial verify pass on all three diffs (PASS; no false-Complete risk). Cross-model codex gate deferred (ChatGPT usage limit) — will run on reset.

🤖 Generated with Claude Code