fix: prevent canonical-csv merge from collapsing distinct conferences

[JesperDramsch]

Root cause analysis

A canonical-csv run (d76e737) renamed PyCon South Africa to PyCon Africa (grafting on africa.pycon.org sponsor/finaid/youtube links) and deleted the real PyCon Africa (Kampala, Uganda) entry.

Why it happened:

1. conference_scorer takes the max of several fuzzy strategies. token_set_ratio returns 100 whenever one name's tokens are a subset of the other's — token_set_ratio("pycon africa", "pycon south africa") == 100 (verified; token_sort is 80, plain ratio 80, partial 74).
2. fuzzy_match treated any score >= EXACT_MATCH_THRESHOLD (100) as an exact match and auto-merged with no confirmation prompt — so the CI-safe "default to no" path in query_yes_no never got a chance to reject it.
3. Both YAML rows ("PyCon Africa" and "PyCon South Africa") then mapped to the same remote CSV row ("PyCon Africa"), got the same title_match index, and were silently collapsed into one chimera row by deduplicate() (bfill().ffill().iloc[0]).

The existing rejections.yml mechanism would have prevented this (it's checked before the exact-match branch), but the pair wasn't listed — this is the same class of bug as the already-listed PyCon Austria / PyCon Australia collision.

Fixes

• utils/tidy_conf/interactive_merge.py
  • A score of 100 alone no longer qualifies as an exact match: the names must be genuinely identical (case/whitespace-insensitive) to auto-merge. Score-100-but-different pairs now go through the fuzzy confirmation path, which defaults to "no" in non-interactive CI.
  • A remote row that already belongs to an identically-named YAML entry (or was already claimed this run) can never be fuzzy-merged by a second conference — this blocks the two-rows-collapse-into-one failure mode entirely, independent of scoring.
• utils/tidy_conf/data/rejections.yml — added PyCon Africa ↔ PyCon South Africa / PyCon ZA / PyConZA as a known false-positive pair.
• utils/check_data_loss.py + canonical-source-update.yml — new workflow step after the merge scripts that fails the run before anything is committed if any (conference, year) entry disappeared from conferences.yml/archive.yml/legacy.yml. Verified it flags the actual d76e737 diff (PyCon South Africa (2026) missing) and passes on identity.
• tests/test_fuzzy_match.py — two regression tests reproducing the incident (with and without the identically-named YAML entry present); both conferences must survive with their own data.

Second bug: Google Maps link detected as Mastodon profile

The TBA-enrichment commit on the same PR (29e9231) wrote mastodon: https://www.google.com/maps/place/Transformatorhuis/@52.386807,... onto the FastAPI Conference entry. The generic Mastodon fallback in utils/enrich_tba.py accepted any URL with /@ in the path — and Google Maps viewports are /@lat,lng,zoom. The fallback now requires the path to be exactly /@handle (or /@user@instance), with regression tests in tests/test_youtube_extraction.py.

Third fix: sub-page links that just repeat the homepage

Upstream sources sometimes fill every URL column with the conference homepage — python-organizers' Proposal URL gave PyCon DE 2027 cfp_link: https://2027.pycon.de/, and main has more (PyCon Hong Kong cfp_link, PyCon Greece / PyCon Italy / PyCon Germany 2026 finaid). A cfp_link/sponsor/finaid identical to the main link carries no information, so the sort pipeline (drop_redundant_link_fields in utils/tidy_conf/links.py, wired into sort_yaml.py) now drops them. URLs count as identical only when they differ at most by scheme, www. prefix, or trailing slash — a different path, subdomain, query, or #anchor is a different pointer and survives (e.g. PyCon SG's sponsor: …/index.html#sponsors). Existing offenders on main get cleaned by the next weekly recurring-sort run; new ones are dropped before every automated commit. Tests in tests/test_redundant_links.py.

Testing

• tests/test_fuzzy_match.py: 18 passed (16 existing + 2 new); test_interactive_merge.py: all pass; test_youtube_extraction.py: 10 passed (8 existing + 2 new); test_redundant_links.py: 13 new tests pass.
• Full merge-related suite run against pip-installed deps: identical pass/fail set to unmodified main (the failures there are pre-existing environment/version artifacts, none touched by this change).
• End-to-end check with the real data files (no mocks): remote "PyCon Africa" vs YAML "PyCon South Africa" now resolves via the rejection list; both conferences survive with correct link/place/dates.

The corresponding data repairs for the open conference-updates PR were pushed separately to auto/conference-updates (53e458a, ec8011a, fba6015). Note: until this PR is merged, a canonical-csv webhook run could still re-introduce the bad merge, since the workflow runs the merge scripts from the accumulator branch's code.

🤖 Generated with Claude Code

https://claude.ai/code/session_019bLVuyt4qdsvSpucA2qvgH