Fix performer Wikipedia-link script + harden nickname matching

[dprodger]

Summary

Fixes the performer Wikipedia-link batch script (backend/scripts/verify_performer_references.py) so it actually runs, makes its bulk output usable, and substantially hardens the shared Wikipedia matcher against false-positive matches for performers whose stored name carries a quoted nickname (e.g. "Brother" Jack McDuff, 'Papa' John DeFrancesco).

Four focused commits:

1. Load .env in the script

The script rolled its own sys.path bootstrap instead of using ScriptBase, so it never loaded .env. db_utils then read empty DB_* vars and fell back to the local Postgres socket (host=None). Added the same dotenv-loading block script_base.py uses.

2. Reduce noise and idle time

For bulk runs over 30k+ performers:

• Demote unchanged/skipped rows to DEBUG so the INFO stream shows only real changes and items needing review.
• Add a progress heartbeat every 50 performers with running counts and an ETA.
• Drop the 1.5s per-performer loop sleep — the Wikipedia and MusicBrainz paths already throttle their own live requests, so it was pure idle time (~13h over 30k rows).

3. Strip quoted nicknames in matching (shared WikipediaSearcher)

Names with a decorative quoted nickname failed lookup two ways: OpenSearch returned an album/redirect instead of the canonical article, and the name comparison scored only a partial (45 < the 50 threshold).

• _strip_nickname() removes paired double quotes and smart single quotes; lone apostrophes (O'Brien, D'Angelo) are untouched. The stripped form is only used when it still has >= 2 tokens, so a bare surname ('Doc' West -> West) is not used — it would fuzzy-match unrelated famous people.
• Verify compares the stripped name on both sides (exact, not partial, match).
• Search queries the canonical form first, then the stored name, merging candidates. Extracted _opensearch(), collapsing duplicated branches and no longer caching transient request failures.

4. Reject non-musician and ambiguous matches

Nickname stripping can yield a common "First Last" that collides with a famous different person ("Captain" Kirk Douglas -> the actor; "Virginia" Joe Jones -> the Fluxus musician).

• Non-musician guard: reject pages whose infobox/lead is an actor/athlete/politician with no music signal (a music term protects genuine musicians).
• Disambiguation-corroboration guard: a (...)-disambiguated title requires a birth/death-year or song match before acceptance.
• Strip hatnotes ("For the musician, see ...") before scoring so cross-references don't leak the other subject's keywords.

Scope note

Commits 3 and 4 touch the shared WikipediaSearcher, so they also improve the core ingestion path (#208) and create_artist.py, not just the batch script.

Verification

Verified end-to-end against the live Wikipedia/production data:

Name	Result
"Brother" Jack McDuff	-> canonical Jack_McDuff
'Papa' John DeFrancesco	-> John_DeFrancesco
Miles Davis (plain)	still valid
'Doc' West, "Bugs" Bower, "Bumps" Myers	no match (was: Kanye West, Kris Bowers, Myers)
"Captain" Kirk Douglas	no match (was: actor Kirk Douglas)
"Virginia" Joe Jones	no match (was: wrong Joe Jones)

Adds 20 offline unit tests (test_wikipedia_nickname.py, test_wikipedia_verify_guards.py) covering _strip_nickname and all the guards via crafted HTML.

Related issues

• Capture performer Wikipedia link during ingestion #208 (capture performer Wikipedia link on ingestion)
• WikipediaSearcher: no retry, and transient failures poison the cache as empty results #209 (WikipediaSearcher retry / cache-poisoning) — partially addressed (no longer caches transient failures)
• verify_performer_references.py bypasses shared MusicBrainz client (no retry/backoff/cache) #210 (script bypasses shared MusicBrainz client)

🤖 Generated with Claude Code