Summary of changes since last commit (a7f5ce7 "Add Stage 1 cross-encoder fast filter for the AI match queue")
All changes below are still UNCOMMITTED as of this writing.

FEATURES
--------
- Legal-form conflict scoring (src/normalize/legal-form.js, new file):
  After candidate matching, compares legal-entity suffixes (Ltd/LLC/Inc/GmbH/AG/etc.)
  between source and target names. Boosts confidence when forms agree (synonyms/
  abbreviations), rejects `direct`-link candidates when forms genuinely conflict
  (e.g. GmbH vs AG). Translation-ambiguous forms (LTD/LLC/INC/CORP/PLC/KK/PTY/PVT)
  are treated as a non-diagnostic cluster to avoid false conflicts on China/Japan
  translation variance.

- tm_xref matching method (src/match/tm-xref.js, new file):
  Cross-references patent applicant names against already-vetted
  trademark_holder -> company matches for a new deterministic bridge.

- Dotted legal-suffix recognition (src/normalize/names.js):
  Suffix stripper now recognizes dotted abbreviations (B.V., N.V., A.G., S.p.A.,
  K.K.), previously invisible to both normalization and legal-form scoring.

- Ambiguous fuzzy-tie guard for patents (src/pipeline/match-patent.js):
  dropAmbiguousFuzzyTies() rejects fuzzy matches with multiple similarly-scored
  candidates; patents dropped from the AI/Ollama review path entirely (raised
  FUZZY_MIN_RATIO 85->92, all AI imports removed).

- Stage1 audit tool (src/pipeline/audit-stage1.js, new file):
  Re-runs Ollama review over stage1_confirmed rows to validate the cross-encoder
  classifier's precision on live data; generalized to support both
  trademark_holder and patent_applicant sources.

- Person-name guard in LEI matching (src/match/lei.js):
  Skips LEI matching for names shaped like personal names (comma-format or
  trailing-initial) to reduce false positives.

BUG FIXES
---------
- Legal-form false conflicts: fixed by introducing the ambiguous-cluster concept
  (was wrongly flagging ~21% of valid matches).

- KFT (Hungarian LLC-equivalent) merged into the LTD synonym family -- was
  wrongly flagged as a conflict against "Ltd".

- Dotted-suffix double bug: (1) no dotted regex variants existed at all,
  (2) even after extraction, dots in the matched token weren't stripped before
  the family lookup -- both fixed.

- Generic-word fuzzy collisions: added wholesale, nutraceuticals, alimentos,
  bebidas, food/foods etc. to FUZZY_STOPWORDS after they caused cross-source
  "magnet" false positives.

- Fuzzy confidence ceiling: capped legal-form-boosted fuzzy matches below the
  "settled" auto-skip threshold (0.94) so they still get human/AI review.

- ai-queue Phase 2 accidental backlog run: corrected build:ai-queue usage/docs
  after an unscoped run queued 812K no-match rows instead of just the intended
  uncertain set.

- Unicode EN-SPACE in patent applicant names: worked around via LIKE pattern
  matching where exact string equality silently failed.

- Multiple manually-rejected false-positive matches (person-name collisions,
  generic-word "magnet" collisions) cleaned up via human_rejected.


SINCE THE feature/visualizer BRANCH MERGED INTO master (PR #3, commit fcab835)
-------------------------------------------------------------------------------
Six commits landed on top of that merge (all on branch stage1-fast-filter,
which starts from master right after the merge; the last one, a7f5ce7, is
the "last commit" referenced above and is not yet merged back to master):

1. 1f58bf0 - Added Contacts entity to the visualizer. UI/UX fixes and
   improvements.
   Adds a new "Contacts" entity end-to-end: schema column, store helper,
   collect-db.js ingestion, api-server.js exposure, and visualizer graph.js
   rendering support (src/db/schema.js, src/db/store.js,
   src/pipeline/collect-db.js, src/server/api-server.js, visualizer/src/graph.js).

2. 6ce1782 - Pop-ups improvement
   Reworked the visualizer's node-details popup (visualizer/src/details.js,
   +148/-.. lines), added a crosshairs-gps icon, and adjusted layout/styling
   (main.js, style.css, index.html).

3. 5ef673f - UI/UX fixes and improvements.
   Broader visualizer UI pass: new icons (close-circle, eraser, filter,
   menu-down, pin/pin-off), reworked filters.js and search.js, and layout/
   style updates across index.html, graph.js, main.js, style.css.

4. 15d619f - Add collapsible Stats and Legend panels with icon buttons
   Stats and Legend panels now collapse to floating icon buttons
   (chart-arc, map-legend SVGs) stacked bottom-left; clicking expands the
   panel with a slide animation, both can be open simultaneously. Also
   added a CHANGELOG.md entry.

5. 4f0df28 - Redirect to login on 401 in client API calls
   visualizer/src/api.js now redirects to the login page whenever a client
   API call comes back 401, instead of failing silently/inconsistently.

6. a7f5ce7 - Add Stage 1 cross-encoder fast filter for the AI match queue
   Adds a fine-tuned cross-encoder (cross-encoder/ms-marco-MiniLM-L-6-v2,
   ONNX export loaded via @huggingface/transformers) that scores
   trademark_holder name pairs locally in milliseconds and auto-approves
   confident ones, skipping the slow Ollama review step entirely for those.
   - training/: dataset extraction, training, and threshold-calibration
     scripts (isolated Python venv, never imported by the Node app)
   - src/match/stage1-classifier.js: loads the ONNX model, exposes
     stage1Score()
   - run-ai-queue.js: opt-in via STAGE1_MODE (off/shadow/live, default
     off) — auto-approve only, no auto-reject (calibration found reject
     isn't safe for this population); new stage1_confirmed match_method
     wired through clearAiMatches, migrateSetLinkType, match-tm.js dedup
     guards, check-legal-forms, export-spreadsheets, and the visualizer
     legend
   Shadow- and live-tested against real queue data before merging; full
   writeup in docs/STAGE1-FAST-FILTER-PLAN.md and STATE_AND_CHANGELOG.md.

Everything in the FEATURES/BUG FIXES sections above happened after a7f5ce7,
on top of this same stage1-fast-filter branch, and remains uncommitted.