Use recurring objects instead of DINOv2 embeddings for annotator-sourced FP sequences

[Chouffe]

Closes #35.

The annotator cohort's identity — "same recurring artefact on the same camera" — no longer gets reconstructed from geometry and DINOv2 distance: it is read from recurring_objects.json, which records it authoritatively. Embed only what we cannot identify. Design, with the decisions recorded: docs/specs/2026-08-14-recurring-object-fp-identity-design.md.

What changes

• embed_fp_for_selection.py skips source: pyro-annotator entries. Their embedding path was degenerate anyway — score-less five-field labels tie best_scoring_frame on every frame, so the crop was whatever directory traversal yielded first (FP selection ranks annotator sequences on a confidence column their labels do not have #29). Per-split embeddings_meta.json now holds only sequences whose identity must be inferred.
• build_fp_yolo_dataset.py pins one background image per recurring object ahead of the clustering (new --recurring-objects arg; the ledger joins the stage deps). The mapping inverts each ledger entry's ingested_folders, restricted to registered pinned folders; a pinned folder missing from the ledger, or claimed twice, is a hard error. Frame pick is deterministic: lexicographically first folder of the object, then highest score / first filename.
• build_sequential_dataset.py keeps pinning every sequence, and drops the now-dead filter that excluded annotator rows from the embedding items.
• Shared helpers (partition_pinned, remaining_quota, new folder_to_recurring_object) move to pyro_dataset.fp.selection.

Verification against a pre-change baseline snapshot

• Embeddings: exactly 50 (train) / 8 (val) / 0 (test) rows dropped; test split bit-identical; remaining pool vectors differ by at most 3.7e-07 (GPU batch-composition noise on 16 of 6,448 rows).
• Equivalence proof: fed the baseline vectors minus annotator rows, the new sequential build reproduces the baseline FP selections byte-for-byte in all three splits — the refactor is behaviour-preserving.
• Real rebuild: sequential differs by exactly one train folder (a same-camera, same-azimuth sibling 76 minutes apart — the measured float noise, not the code). YOLO quotas exact (two_stage+50_pinned / two_stage+8_pinned), all 58 pinned objects present, test split unchanged; train/val clustered remainder re-rolls as expected since k changed.
• test_data_leakage: 17 passed. Full suite: 156 passed.

Merge with #36 (append-only test growth)

Main merged the test-growth implementation mid-review; this branch merges it back in (71f3da6) and the two designs compose:

• build_sequential_dataset.py resolves to Grow the sequential test set from annotator imports, append-only #36's structure — the test split copies the frozen lockfile verbatim, no selection — with this PR's train/val simplifications inside the else-branch (helpers imported from pyro_dataset.fp.selection, dead pinned_folders filter gone).
• The one semantic tension between the two specs resolves in this PR's favour by construction: freeze_test_selection.py pins annotator test FPs straight from the registry (no embeddings needed), and its clustering fill simply cannot see un-embedded surplus pins — they defer until new test smoke opens slots, instead of competing by embedding distance. Recorded in the spec; the test-growth spec's §5 wording ("surplus eligible for two-stage selection") is weakened accordingly.
• Grow the sequential test set from annotator imports, append-only #36's four new frozen_test_selection tests are kept in tests/test_build_sequential_pinning.py; the pinning-helper tests live in tests/test_fp_pinning.py where this PR moved them.
• Doc staleness the merge created is fixed: the spec no longer claims annotator sequences cannot enter test (that refusal was repealed by Grow the sequential test set from annotator imports, append-only #36), and CLAUDE.md's pinning bullet is scoped to train/val with the freeze-quota path described for test.

Post-merge verification (all local, real stages):

• Full dvc repro on the merged tree: frozen test negatives materialise exactly (built sequential_test FP set == lockfile, 151/151); leakage gate passes with Grow the sequential test set from annotator imports, append-only #36's new recurring-object-in-test assertions (19 tests); dvc status clean; 189 unit tests pass.
• Determinism byte-match: re-running the merged builder into a scratch directory reproduces both committed sequential datasets byte-for-byte (diff -qr across 3,208 train + 330 val + 302 test sequence folders, images and labels — zero differences).

Notes

• Data for the rebuilt outputs is pushed to the DVC remote (229 objects — content-addressed dedup keeps the delta small).