Hard-negative probe retrain on the shared Pillow path

[felirami]

What

Retrains the DINO probe head on features from the shared Pillow-exact resize, with 1,915 hard-negative reals added to training, and re-derives the fusion rescue bands for the cleaner probe. Three commits:

1. DINO preprocess unification (src/dino.js, eval/extract-features.mjs): extension, Node harness, and probe training now compute identical DINO inputs through src/pixel-resize.js. The extractor's private preprocess duplicate is deleted (it had drifted and carried an unguarded grayscale read). Browser-vs-Node DINO parity measured in a live browser: max delta 0.000046, down from 0.22.
2. fetch-train hardening: resumable (rate-limit aborts resume cheaply), authenticated (HF token picked up automatically), and extended with the five hard-negative real sources (Unsplash stock, Amazon Berkeley product shots, LSUN bedrooms, DeepFashion catalog, Oxford flowers; rows strictly disjoint from the eval stress set).
3. Probe + bands (models/probe/dino-probe.json, src/fuse.js): retrained on 11,405 rows; bands re-derived under the stress guard with a conservative near-optimum chosen over the grid maximum.

Why hard negatives

The naive Pillow-path retrain regressed on professional stock photography: the training reals (COCO, Flickr, faces, food, scenes) contain no polished stock/product distribution, so the probe saturated on exactly the images that sank claim 1014-style submissions. Adding 1,915 such reals to training removed every DINO-attributable stress false positive.

Measured results (harness runs, not simulation)

893-image public bench at raw 0.65:

Configuration	BA	TPR	TNR
v1.1.0 (shipped)	87.7%	75.8%	99.6%
Naive Pillow retrain (not shipped)	87.2%	74.8%	99.6%
This PR	90.5%	81.2%	99.8%

Per-generator recall: DALL-E 3 93.3% approx, MJHQ 96.7%, Flux 1.1 83.3%, GPT-4o image 60.0%, SD-ELSA 78.5% (winner-detail figures; the shipped conservative variant is within about a point of each).

240-image full-resolution stock/catalog/product stress set, identical bytes across configurations:

Configuration	Stress FPs	Of which DINO-attributable
v1.1.0 (shipped)	7	3
Naive Pillow retrain	9	5
This PR	4	0

The 4 remaining stress FPs are CommunityForensics alone scoring >= 0.65 on full-resolution Unsplash originals, where CF is authoritative by design. Worth checking exactly that during the pre-claim live smoke.

Full suite: 213 tests green. Constants chosen conservatively: DINO_RESCUE_MIN stays 0.70 and the CF-hard-zero guard stays, giving up 1.4 bench BA versus the grid maximum in exchange for margin on unmeasured live distributions.

Caveats

• The hard-negative additions to fetch-train mirror the exact datasets-server rows and rules that built the shipped probe's negatives, but an end-to-end refetch validation is pending a rate-limit cooldown (today's fetch volume exhausted even the authenticated tier).
• Public fixture numbers are directional and are not a claim about the private benchmark. Live-site smoke remains required before any claim.

---

Note

Medium Risk
Changes core scoring (probe weights and fusion constants) and DINO preprocessing parity; bench/stress metrics are documented but live-site smoke and full refetch validation are still caveated in the PR.

Overview
Retrains and ships a new DINO logistic probe (models/probe/dino-probe.json, ~11.4k training rows) on features from the shared Pillow-exact resize, with ~1.9k hard-negative reals (stock, catalog, product, interiors) added via extended eval/fetch-train.mjs (resumable fetches, optional HF_TOKEN, disjoint stress-set rows).

DINO inputs are unified end-to-end: src/dino.js and eval/extract-features.mjs now use pillowResize + cropPackedPixels (optional cropSize in clip-preprocess.js) so extension, Node harness, and training match; the extractor’s duplicate preprocess is removed.

Fusion rescue tiers are re-derived in src/fuse.js for the new probe (e.g. strong band 0.02–0.10 / DINO ≥ 0.90, sub-floor CF ≥ 0.0005 / DINO ≥ 0.995). Docs and tests reflect harness-reported 90.5% BA on the 893-image fixture and 0 DINO-attributable stress-set FPs vs v1.1.0’s 7 total / 3 DINO-driven on the same bytes.

^{Reviewed by Cursor Bugbot for commit 5fcaba9. Bugbot is set up for automated code reviews on this repo. Configure here.}