feat(vlm): optional per-pack registry prior for open-vocabulary labeling (#61)

[davidamacey]

Refs #61 (item 2). Not closing: the acceptance (overall accuracy on the COCO oracle, prior on vs off) is a live measurement that is still pending.

What

A prompt pack gets registry_prior_top_k (integer, 0 = off by default, hard max 50). When > 0, open-vocabulary VLM labeling (POST /vlm/label_batch and the auto-label pipeline VLM stage) adds to the prompt the top-k registry class names, ranked by validated-item count (ties by name), plus up to k pending new-class proposal names.

Design

• Built server-side from project state (registry_prior_source.py, aggregations on the bound project's items index, holdout excluded); no client-supplied lists. Names only, never indexes.
• A hint, not a constraint: reply parsing is unchanged and still accepts any answer.
• Fail closed: if the pack asks for the prior and the counts cannot be read, label_batch returns 503 and the pipeline returns a stage error, rather than labeling without it.
• Validation: non-integer or bool is pack_field_missing; outside 0..50 is the new pack_field_out_of_range. Served in the pack schema as kind: "int", group open_classify. Contracts regenerated; frontend PackSchemaField.kind, PackFieldValue, fixture, test pin and CHANGELOG updated (the editor shows an unknown kind read-only and saves it unchanged).
• Lock rule and test_holdout untouched: only prompt text changes.
• Not done: a per-item prior from detector_class_name; this is a per-batch registry prior.

Measurement harness (offline-tested only)

scripts/curation/bakeoff/vlm_prior_oracle.py (dump per run, compare off vs on). No live VLM/GPU run was done.

Verified

Items 1 (proposal denylist) and 3 (label_agreement) of the issue are already on main.