Give wordless pictures a way to be found

A screenshot with no OCR-able words used to be a memory nothing could
reach. It now has a vector, and a question can be asked with it.

- picture space: Xenova/chinese-clip-vit-base-patch16 fp16 ONNX, 361MB,
  sha256 0fad978faa89b85f…, CPU-runnable on the onnxruntime the repo
  already ships. Text 76ms, image 29ms/image at batch 8. The ONNX contract
  was measured, not assumed: it is the fused contrastive graph, so the
  unused tower is fed a placeholder row, and the tokenizer must be
  BertTokenizer — a CLIP tokenizer emits ids past the embedding table.
- visual_index.py keeps picture vectors in their own family beside the
  text ones, keyed by atom id, and refuses to answer when the map and the
  array disagree. retrieval.py stays under 500 lines.
- a search can be asked with a picture, and when the words come up thin
  the picture ranking joins the fusion — but only then, so a text query
  never pays for the encoder.
- visual tokens are counted at 32x32 patches (migration v7): 21,325 image
  atoms went from 5,331 to 34,618 tokens, and the library from 29.9M to
  84.2M. The 100M capsule fits this 50,278-file corpus and nothing much
  larger — that is what the claim actually means.
- the vision-language describer is written and gated: with no snapshot an
  atom records "deferred: no vision-language model snapshot"; a broken
  model becomes a note on the atom instead of a dead ingest.
- forget now voids the picture index and the copy this process holds, so
  a forgotten picture stops answering from RAM.

Measured, six arms, ScreenSpot-v2 (554 icon crops, 408 whole screenshots),
bench/reports/VISUAL.md:
  icon pool      picture 0.208  words 0.175  chance 0.018
  screenshot pool words 0.384   picture 0.215
  screenshot to icon 0.108  — CLIP is a whole-image model, not a grounding
  model: it cannot point at one small icon inside a screen.
  within one screen MRR 0.908 (recall@10 is 1.000 there only because
  chance is 1.000 too, so MRR is the only honest number in that row).
ScreenSpot-v2 instructions are all English, so this run does not test the
bilingual claim and the report says so.

Gates: 199 tests, ruff clean on 17 files, mypy clean on 16 (fresh cache),
the repo's own pyright 0 errors on the 12 files touched.

CHECKLIST.md

@@ -134,11 +134,24 @@
 
 ## 阶段 2 — P1：让它成为产品
 
-- [ ] **p1-1** 图像进 dense
-  - [ ] p1-1.1 选 embedding ≤500MB，双语 + 图像；记录模型 SHA 与体积
-  - [ ] p1-1.2 图像 atom 入索引；query 侧支持图像
-  - [ ] p1-1.3 视觉 token 计数改 32×32 口径（1280×720 ≈ 900）
-  - [ ] p1-1.4 VLM 描述：代码路径就绪，标 `DEFERRED`
+- [x] **p1-1** 图像进 dense
+  - [x] p1-1.1 选 embedding ≤500MB，双语 + 图像；记录模型 SHA 与体积
+    - `Xenova/chinese-clip-vit-base-patch16` fp16 ONNX，361MB，`onnx/model_fp16.onnx` sha256 `0fad978faa89b85fda53a83a4c202116a9eb87c8f9723fb4d85e4483efbce4b2`；CPU 可跑（仓库已有 onnxruntime）；文本塔 76ms、图像塔 29ms/张（batch 8）
+    - 契约是量出来的不是猜的：输入 `input_ids`/`attention_mask`/`pixel_values (N,3,224,224)`，输出 `image_embeds`/`text_embeds (N,512)`；这是融合对比图，未用的塔必须喂占位行；分词器用 `BertTokenizer`（按 CLIP 分词器加载会得到超出 vocab 的 id，Gather 直接炸）
+  - [x] p1-1.2 图像 atom 入索引；query 侧支持图像
+    - `backend/src/ely_eye/visual_index.py`（独立文件，retrieval.py 停在 499 行没有膨胀）；`ely-eye index-visual`；`search --image`；API `SearchRequest.image`
+    - 开关规则：`clip_model_dir` 指向含 `onnx/model_fp16.onnx` 的目录才算可用，不再多一个 bool
+    - 14 个测试：句子找到图、图找到图、新图一到索引即失效、缺快照点名、map/vector 不一致点名、词够多时不付 encoder、forget 之后图片不再从内存回答
+  - [x] p1-1.3 视觉 token 计数改 32×32 口径（1280×720 ≈ 900）
+    - `visual_token_equivalent = ⌈w/32⌉·⌈h/32⌉`；migration v7 重算存量：21,325 个图像 atom 从 5,331 → 34,618 tokens；整库 29.9M → **84.2M**，于是「100M 胶囊」在这份 50,278 文件语料上刚好装得下、再大就装不下——这才是那句 claim 的真实含义
+  - [x] p1-1.4 VLM 描述：代码路径就绪，标 `DEFERRED`
+    - `describe.py`；ingestion 只在 OCR 没找到字时才调用（图上已有字是免费的）；无快照 → atom metadata `deferred: no vision-language model snapshot`；模型报错 → `failed: …` 且 ingest 不死（策略在调用方，不在 describer）；`status` 显示两个视觉开关；5 个测试
+- p1-1 证据（我额外加的六臂基准，不是验收项）：`bench/visual_recall.py`，ScreenSpot-v2 的 554 个图标 crop + 408 张整屏，报告 `bench/reports/VISUAL.md`
+  - 图标池：picture 0.208 vs words 0.175 vs chance 0.018 —— 图片空间比瞎猜高 11.5 倍，并且在 427 个 OCR 没字的 crop 上超过了词
+  - 整屏池：words 0.384 > picture 0.215 —— 整屏的身份主要是文字，OCR 词在这里赢
+  - 整屏 → 图标：0.108 —— CLIP 不会在截图里定位小图标，它是整图相似模型，不是 grounding 模型
+  - 单屏内：MRR 0.908（recall@10 是 1.000 但 chance 也是 1.000，所以那一行只有 MRR 有意义）
+  - 本轮没测双语：ScreenSpot-v2 的指令全是英文，报告里写明了
   - 验收：无文字图像被图像查询召回有 recall 数字；`pytest tests/test_visual_tokens.py` 绿
 - [ ] **p1-2** 音频摄入
   - [ ] p1-2.1 `Modality` 加 audio/transcript；未知扩展名不再静默跳过