feat(backend): scan-product accuracy 66.2% -> 79.7% + frozen validation benchmark

Accuracy work on the 79-image product-scan validation set (user goal: 90%):
- classify_ocr_server.py: 0/90/180/270-degree expiry-date search (stops at
  first hit, 0-degree fallback); classification decoupled onto the upright
  image (rotated frames regressed DINOv2 -6pts until this); cross-line date
  stitching; tiled full-res OCR pass (defeats the 4000px downscale that
  killed small inkjet dates); VL-pipeline expiry fallback with
  keyword-anchored anti-hallucination guard; VL text lines merged into
  text_lines + VL SKU retry. Visualization endpoints removed entirely
  (Visual/Spotting grids - unused by frontend, 3x per-scan GPU cost).
- product-scan.ts: coverage-normalized OCR-evidence re-ranking of DINOv2
  top-K (tuned offline: +8/-0 on top-1 misses), re-ranked class mapped to
  sku_master by SKU prefix; classifier timeout 90s->240s for fallback paths.
- Frozen benchmark: product-test-images-fixed/ (79 renamed images) +
  freeze/seed/build-undetected/capture/experiment scripts; labels trimmed to
  the 79 validation entries (training rows kept in .bak-with-training);
  5 TRAINED-ON SKUs replaced with fresh held-out photos.
- manual-label-scan page: shows last batch-test AI prediction under every
  field by default (new /api/product-scan-results); serves the fixed folder;
  fixed total hydration failure via allowedDevOrigins 127.0.0.1.
- Measured (all-79, zero failures): sku/name 87.3%, expiry 64.6%, overall
  79.7%. Tiles/VL-evidence/VL-SKU deployed but not yet batch-measured.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gr6HH7JrdsXX8AARejQboM

backend/sources/product-test-images-fixed/README.md

@@ -0,0 +1,29 @@
+# Product-scan validation images — frozen benchmark set
+
+This is the **actual Validation Set the accuracy harness scores**
+(`accuracy-check-scan.mts`'s `getImagePath()` points here, not at
+`../product-test-images/`). It exists so re-running the harness always grades
+the exact same images — the live-intake folder can keep growing from new
+`/manual-label-scan` drops without silently shifting the benchmark underfoot.
+
+**Naming**: each file is `<index> <no_sku>.<ext>` (e.g. `1 11110059.jpeg`),
+where `<index>` is just this file's stable position in the set — it carries
+no other meaning. `product_manual_labels.json`'s ground-truth entries for
+these images use this same filename.
+
+**Do not hand-edit this folder.** It's fully generated by
+`node scripts/freeze-validation-set.mjs` (run from `backend/`), which copies
+every flat (non-training) entry out of `product_manual_labels.json` from
+`../product-test-images/`, renames it, and rewrites those entries'
+`filename` fields to match. To add a new SKU/photo to the benchmark:
+label it in the live-intake folder first (see that folder's README), then
+re-run the freeze script.
+
+**79 images as of 2026-07-14.** 5 of them (SKUs 12010801, 12012504, 12130504,
+13050101, 15040102) are flagged `TRAINED-ON` in `product_manual_labels.json`'s
+`notes` field — their only available source photo (from an external
+`research-sam3` segmentation project) was already used to train the
+classifier (as a SAM3 crop + augmentations), so they are **not** a clean
+held-out test. Their per-image scores will read as memorization, not real
+generalization, until a fresh, never-trained-on photo is dropped for those
+SKUs. The other 74 are genuinely held out.