Dw Color Image Audit

Second mapper-completeness gap the dash-fix (307493e) didn't cover. On NO-DASH
titles the whole title is scanned and a botanical/place/motif homonym in the
lexicon (grass->green [grasscloth], ocean->teal, marine->blue ['Cote Marine'],
leaf->green, apple->chartreuse ['Big Apple'], forest->green ['Fir/Virgin Forest'])
was mis-read as the colorway, inflating raw-HIGH noise.

Fix (DTD verdict A, 3/3 Claude+Codex+Qwen): add AMBIGUOUS_COLORWORDS set +
demote_ambiguous flag in _familyFromPhrase / _family_from_phrase. Demotion is
gated to the NO-DASH / whole-title path ONLY — the clean after-dash colorway
segment is untouched, so 'X - Ocean' still resolves teal and Hawksbill-Forest /
Hawksbill-Ocean stay genuine HIGHs. When a real colorway co-occurs it survives
and wins; when the only signal is a homonym, fall through to UNKNOWN instead of
asserting a wrong family. Mirrored identically in JS + both Python mappers.

Read-only rerun on the existing flagged set (cached hexes, no image re-download,
$0, no network): vs the post-dash-fix baseline HIGH 629->602 (27 false-positive
HIGHs cleared), 26 homonym-only no-dash titles routed HIGH->UNKNOWN, ZERO new
HIGH regressions (new_high held at 21, all dash-fix unmaskings). JS<->Python
parity verified across all 1831 titles (0 real mismatches). Self-test 13/13 PASS.
No DB/Shopify writes; catalog disposition unchanged (~925 set stays HOLD).

Root cause: nameToFamily/name_family kept the LAST mappable token in a DW title's
color segment. DW titles read 'Pattern - Colorway MaterialType' so the trailing
'Vegan Leather' (leather->brown in lexicon) overrode the true colorway. 27 Tier-1
HIGH rows mis-bucketed brown.

Fix (DTD verdict A, 3/3 Claude+Codex+Qwen): dash-aware colorway parse (segment
after last ' - '), material-stopword skip gated on a real color surviving,
multi-word colorway map (sea green->teal), fallback to whole segment when the
after-dash part has no color (preserves pattern-half colorways). Same fix mirrored
in JS (color-classify.js) and both Python mappers (reextract.py/reextract2.py).
Added reextract2 __main__ guard so the read-only rerun harness imports the real code.

Re-run on existing flagged set (cached hexes, no image re-download, $0): 88
declaredFamily corrected, 17 false-positive HIGHs cleared (incl. Baroque-Orange,
Pinpoint-Sea Green, Crushed Coral-Amber), 21 previously-MASKED real mismatches
unmasked (Sapphire/Azure/Lagoon/Peacock... on off-hue images). Real alarm preserved
(Hawksbill-Walnut brown vs orange still HIGH). No DB/Shopify writes.