feat: fine-tune the identity head, and measure what it buys

Part 2 ended on the number that matters: enrolling a real shelf from itself leaves
off-the-shelf DINOv2 at ~25% top-1, because real milk cartons differ by small text
and a colour band. This trains a small projection head on frozen features - 384 -> 128,
cosine classifier, angular margin - with augmentation standing in for the second visit
we do not have, and measures it on 72 held-out facings over 8 products.

  raw DINOv2 cosine               top-1  2/72 (2%)   balanced  1%
  raw, rejecting below 0.55       top-1  2/72 (2%)   balanced  1%   (5% of answered right)
  fine-tuned + margin             top-1 28/72 (38%)  balanced 73%
  fine-tuned, rejecting below 0.4 top-1 17/72 (23%)  balanced 58%  (94% of answered right)

So fine-tuning is the difference between working and not (balanced 1% -> 73%), and
with a rejection threshold the model becomes precise - it answers 18 of 72 facings and
is right on 94% of them, which is the operating point a store wants. The raw model
cannot do that at any threshold.

Where it fails is where my labels are coarse: MEGMILK, which the contact sheet shows
is probably several products under one brand, scores 1/32 because the class is
internally inconsistent and the model correctly refuses most of it.

Honest limit, stated in the doc: enrolment and held-out facings come from one
photograph, so these are upper bounds; only a recorded second visit settles the real
number. What is settled is the mechanism - a head trained on labels the pipeline can
produce itself moves fine-grained identity from useless to usable.

175 tests pass.

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>

recognization/docs/EXTERNAL-EVAL.md

@@ -180,3 +180,55 @@ So the honest design is: **reject by default**, name the product from the tag wh
 is readable, from the packaging text when it is legible, from a fine-tuned embedding
 otherwise - and hand the remainder to the operator, whose answers become the next
 enrollment.
+
+
+---
+
+# Part 3: fine-tuning the embedding, and what it buys
+
+Part 2 ended on the number that matters: enrolling a real shelf from itself leaves
+off-the-shelf DINOv2 at ~25 % top-1, because real milk cartons differ by small text
+and a colour band. The literature's answer is fine-tuning, so this measures it.
+
+`code/finetune_identity.py` trains a small projection head (384 -> 128, cosine
+classifier, angular margin) on frozen DINOv2 features, with augmentation standing in
+for the second visit we do not have. Two enrolment views per product, twelve
+augmented copies each, 72 held-out facings, 8 products. Balanced accuracy is reported
+because the shelf is unbalanced (MEGMILK has 34 facings, 酪農3.6 has two - a system
+that always answered "MEGMILK" would look respectable on the raw rate).
+
+| | top-1 | balanced | answered | of those right |
+| --- | --- | --- | --- | --- |
+| raw DINOv2 cosine | 2/72 (2 %) | 1 % | 72 | 3 % |
+| raw, rejecting below 0.55 | 2/72 (2 %) | 1 % | 44 | 5 % |
+| **fine-tuned + margin** | 28/72 (38 %) | **73 %** | 72 | 39 % |
+| **fine-tuned, rejecting below 0.4** | 17/72 (23 %) | 58 % | 18 | **94 %** |
+
+Three things follow.
+
+1. **Fine-tuning is not a refinement, it is the difference between working and not.**
+   Balanced accuracy goes from **1 % to 73 %** on the same held-out facings. Off-the-
+   shelf features cannot separate these products; a head trained on sixteen crops can.
+2. **With a rejection threshold it becomes precise**: answering only 18 of 72 facings
+   and being right on 94 % of them is exactly the operating point a store wants -
+   name what you are sure of, and hand the rest to the operator. The raw model cannot
+   do this at any threshold (5 % of answered).
+3. **Where it fails is where the labels are coarse.** Per class, with rejection:
+   明治ラブ 2/2, おいしい無調整豆乳 2/2, 低脂肪 1/1, 特濃4.5 1/2, 北海道牛乳 8/21,
+   明治 おいしい牛乳 2/12, and **MEGMILK 1/32**. MEGMILK is the class I labelled most
+   coarsely - the contact sheet shows what are probably several different products
+   under one brand - so the class is internally inconsistent and the model correctly
+   refuses most of it. Better labels, not a better model, is what that needs.
+
+## The honest limit
+
+Enrolment and held-out facings come from **one photograph**. Augmentation stands in
+for a second visit, which is not the same thing: a real second visit changes the
+light, the angle and the stock. So 73 % balanced and 94 % precision are **upper
+bounds**, and the only way to know the real number is to record a second visit
+(`docs/DATA.md`).
+
+What is settled is the mechanism: **a head trained on labels the pipeline can produce
+itself** - tag-named facings, plus operator confirmations - moves fine-grained
+identity from useless to usable, and a rejection threshold turns it into something a
+store can act on.