Extract stage: vision title + edition-cue extraction, offline-tested

Per-photo raw results cached under data/extract_raw/ (gitignored) so
re-runs are free, --only re-extracts a single photo, and titles.json is
rebuilt with dedupe that keeps conflicting-edition sightings separate.
HEIC converts via pillow-heif; images downscale to <=1568px long edge;
model JSON parsed defensively (code fences, surrounding prose). Vision
callable is injectable — tests use a local fake; the real one uses the
anthropic SDK (approved) with the model from config.toml.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

pyproject.toml

@@ -8,6 +8,10 @@ dependencies = [
     "httpx>=0.27",
     "rapidfuzz>=3.9",
     "defusedxml>=0.7.1",
+    "anthropic>=0.120.2",
+    "pillow>=12.3.0",
+    "pillow-heif>=1.5.0",
+    "rich>=15.0.0",
 ]
 
 [project.scripts]