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

CLAUDE.md

@@ -145,9 +145,13 @@ What differs from the platform loop:
   the next `dvc repro`, or the build errors on the stale lockfile.
   Design: `docs/specs/2026-08-14-annotator-test-growth-design.md`.
 - **False-positive boxes are class 99**, not `0`: they mark where the detector
-  fired, not smoke. `build_sequential_dataset.py` pins annotator-sourced FP
-  sequences ahead of its clustering, so the ones chosen upstream actually reach
-  the built dataset.
+  fired, not smoke. Annotator-sourced FP sequences are never embedded — their
+  identity comes from `recurring_objects.json` — and in train and val both
+  builders pin them ahead of clustering: the sequential build takes every
+  sequence, the YOLO build one image per recurring object. In test they reach
+  `sequential_test` only through `freeze_test_selection.py`, up to the quota —
+  surplus pins are deferred until new test smoke opens slots. See
+  `docs/specs/2026-08-14-recurring-object-fp-identity-design.md`.
 - Each folder carries a `meta.json` with every lane's full track, including boxes
   left out of `labels/`, so an object-level dataset can be derived later without
   re-exporting.

data/reporting/leakage_results.xml

@@ -1 +1 @@
-<?xml version="1.0" encoding="utf-8"?><testsuites><testsuite name="pytest" errors="0" failures="0" skipped="0" tests="17" time="0.315" timestamp="2026-08-14T12:46:10.441113+02:00" hostname="desktop"><testcase classname="tests.test_data_leakage" name="test_wildfire_yolo_no_image_leakage[train-test]" time="0.018" /><testcase classname="tests.test_data_leakage" name="test_wildfire_yolo_no_image_leakage[val-test]" time="0.003" /><testcase classname="tests.test_data_leakage" name="test_wildfire_yolo_no_image_leakage[train-val]" time="0.017" /><testcase classname="tests.test_data_leakage" name="test_yolo_datasets_no_image_leakage[train_val_dir0-train]" time="0.020" /><testcase classname="tests.test_data_leakage" name="test_yolo_datasets_no_image_leakage[train_val_dir1-val]" time="0.005" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a0-dir_b0-train-val-wildfire]" time="0.005" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a0-dir_b0-train-val-fp]" time="0.005" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a1-dir_b1-train-test-wildfire]" time="0.005" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a1-dir_b1-train-test-fp]" time="0.005" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a2-dir_b2-val-test-wildfire]" time="0.001" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a2-dir_b2-val-test-fp]" time="0.001" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a0-dir_b0-train-val-wildfire]" time="0.062" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a0-dir_b0-train-val-fp]" time="0.038" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a1-dir_b1-train-test-wildfire]" time="0.063" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a1-dir_b1-train-test-fp]" time="0.037" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a2-dir_b2-val-test-wildfire]" time="0.012" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a2-dir_b2-val-test-fp]" time="0.007" /></testsuite></testsuites>
+<?xml version="1.0" encoding="utf-8"?><testsuites><testsuite name="pytest" errors="0" failures="0" skipped="0" tests="19" time="0.538" timestamp="2026-08-14T16:28:20.043266+02:00" hostname="desktop"><testcase classname="tests.test_data_leakage" name="test_wildfire_yolo_no_image_leakage[train-test]" time="0.032" /><testcase classname="tests.test_data_leakage" name="test_wildfire_yolo_no_image_leakage[val-test]" time="0.006" /><testcase classname="tests.test_data_leakage" name="test_wildfire_yolo_no_image_leakage[train-val]" time="0.031" /><testcase classname="tests.test_data_leakage" name="test_yolo_datasets_no_image_leakage[train_val_dir0-train]" time="0.036" /><testcase classname="tests.test_data_leakage" name="test_yolo_datasets_no_image_leakage[train_val_dir1-val]" time="0.008" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a0-dir_b0-train-val-wildfire]" time="0.006" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a0-dir_b0-train-val-fp]" time="0.008" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a1-dir_b1-train-test-wildfire]" time="0.008" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a1-dir_b1-train-test-fp]" time="0.009" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a2-dir_b2-val-test-wildfire]" time="0.002" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_sequence_leakage[dir_a2-dir_b2-val-test-fp]" time="0.002" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a0-dir_b0-train-val-wildfire]" time="0.101" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a0-dir_b0-train-val-fp]" time="0.061" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a1-dir_b1-train-test-wildfire]" time="0.097" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a1-dir_b1-train-test-fp]" time="0.059" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a2-dir_b2-val-test-wildfire]" time="0.017" /><testcase classname="tests.test_data_leakage" name="test_sequential_no_image_leakage[dir_a2-dir_b2-val-test-fp]" time="0.010" /><testcase classname="tests.test_data_leakage" name="test_no_recurring_object_spans_splits" time="0.005" /><testcase classname="tests.test_data_leakage" name="test_lockfile_matches_the_test_quota" time="0.005" /></testsuite></testsuites>

docs/specs/2026-08-13-annotator-export-sequential-import-design.md

@@ -434,6 +434,10 @@ This does not displace anything. The import adds as many smoke sequences as the
 holds (58 here), which raises `quota = n_wf` by the same number — the pinned FPs consume
 exactly the slots their own positives created.
 
+Since [2026-08-14-recurring-object-fp-identity-design.md](2026-08-14-recurring-object-fp-identity-design.md),
+pinning extends to the YOLO build (one image per recurring object) and the
+cohort is no longer embedded at all.
+
 ### 7. Changes to existing pyro-dataset code
 
 Everything else is new code. Exactly two existing files change, and both changes are

docs/specs/2026-08-14-recurring-object-fp-identity-design.md

@@ -136,8 +136,11 @@ score, ties broken by filename. Annotator labels are score-less, so in practice
 the first labeled frame by name — but the rule stays correct if a scored sequence ever
 carries a `source`.
 
-Test split: untouched. Annotator sequences never enter test
-(`assignments_from_splits` refuses it), so round-robin is unaffected.
+Test split: untouched by this design. When it was written, annotator sequences
+could not enter test (`assignments_from_splits` refused it); the test-growth
+design has since repealed that refusal, so annotator test FPs will eventually
+appear in the round-robin pool — accepted, as the test-growth design declares
+detector-test behaviour out of its scope.
 
 ### `scripts/build_sequential_dataset.py`
 
@@ -196,8 +199,12 @@ Integration verification, in order:
   beyond fewer sequences passing through KMeans.
 - Interaction with
   [2026-08-14-annotator-test-growth-design.md](2026-08-14-annotator-test-growth-design.md)
-  (designed separately): it opens `split: test` to annotator imports, repealing the
-  "test is never assigned" premise this document leans on, and its §5 makes surplus
-  annotator test FPs "eligible for two-stage selection" — which presumes embeddings
-  this design stops producing. That reconciliation belongs to the test-growth
-  implementation, not to this one.
+  (designed separately, merged since): it opens `split: test` to annotator imports,
+  repealing the "test is never assigned" premise this document originally leaned on.
+  Its §5 makes surplus annotator test FPs "eligible for two-stage selection", which
+  presumes embeddings this design stops producing. How that lands in practice:
+  `freeze_test_selection.py` pins annotator test FPs from the registry — no
+  embeddings needed — and its clustering fill simply cannot see the surplus, so a
+  deferred sequence waits for future slots instead of competing by embedding
+  distance. Weaker than §5's letter, consistent with both designs' spirit: exact
+  identity is never traded back for estimated identity.

dvc.lock

@@ -592,18 +592,18 @@ stages:
   build_fp_yolo_dataset:
     cmd: uv run python scripts/build_fp_yolo_dataset.py --registry data/raw/fp/registry.json
       --data-dir data/raw/fp/data --wf-dataset data/processed/wildfire_yolo --embeddings-dir
-      data/interim/fp_sequence_embeddings --output data/processed/fp_yolo --random-seed
-      0
+      data/interim/fp_sequence_embeddings --recurring-objects data/raw/pyro-annotator/recurring_objects.json
+      --output data/processed/fp_yolo --random-seed 0
     deps:
     - path: data/interim/fp_sequence_embeddings/train
       hash: md5
-      md5: 51edec8cc1a3f1c8b93e7af595f80717.dir
-      size: 20216589
+      md5: d7c1ce9531de9fc9d89bf6f4305ba12d.dir
+      size: 20043098
       nfiles: 2
     - path: data/interim/fp_sequence_embeddings/val
       hash: md5
-      md5: aac0d5327b7b6ea3e05a66dccf154749.dir
-      size: 2491101
+      md5: 35a4487a1ba6a41ca2b529c6a699e6c3.dir
+      size: 2463314
       nfiles: 2
     - path: data/processed/wildfire_yolo
       hash: md5
@@ -619,19 +619,23 @@ stages:
       hash: md5
       md5: 7983fceca0836f734bdf7e6bf032063e
       size: 1318880
+    - path: data/raw/pyro-annotator/recurring_objects.json
+      hash: md5
+      md5: ecc9988f92c56db6f9516052b8aae959
+      size: 50832
     - path: scripts/build_fp_yolo_dataset.py
       hash: md5
-      md5: ca087d24c82e3a93ff766c015e852a37
-      size: 10573
+      md5: 7017e8e7f6d9d2d9e1ca981b63f25e7c
+      size: 13137
     - path: src/pyro_dataset/fp/selection.py
       hash: md5
-      md5: e9d2065653a9aafdc035c700fa2d678f
-      size: 7999
+      md5: e8b6b6673a571baeba4f94e18faeb98a
+      size: 10121
     outs:
     - path: data/processed/fp_yolo
       hash: md5
-      md5: cf82b8283d478d11ae8458335327f778.dir
-      size: 332638599
+      md5: 218aa610b9887a65ec9b27009d6a6f6f.dir
+      size: 332994040
       nfiles: 6002
   merge_yolo_dataset:
     cmd: uv run python scripts/merge_yolo_dataset.py --wf-dataset data/processed/wildfire_yolo
@@ -640,8 +644,8 @@ stages:
     deps:
     - path: data/processed/fp_yolo
       hash: md5
-      md5: cf82b8283d478d11ae8458335327f778.dir
-      size: 332638599
+      md5: 218aa610b9887a65ec9b27009d6a6f6f.dir
+      size: 332994040
       nfiles: 6002
     - path: data/processed/wildfire_yolo
       hash: md5
@@ -660,17 +664,17 @@ stages:
       nfiles: 5281
     - path: data/processed/yolo_train_val
       hash: md5
-      md5: 038270d9336432fa9001cf92069731b2.dir
-      size: 2027739474
+      md5: a43f8eae91dd781aa91984d92bbe65d2.dir
+      size: 2028094915
       nfiles: 33635
   visualize_yolo_train_val:
     cmd: uv run python scripts/visualize_dataset.py --dir-dataset data/processed/yolo_train_val
       --dir-save data/reporting/viz/yolo_train_val
     deps:
     - path: data/processed/yolo_train_val
       hash: md5
-      md5: 038270d9336432fa9001cf92069731b2.dir
-      size: 2027739474
+      md5: a43f8eae91dd781aa91984d92bbe65d2.dir
+      size: 2028094915
       nfiles: 33635
     - path: scripts/visualize_dataset.py
       hash: md5
@@ -679,8 +683,8 @@ stages:
     outs:
     - path: data/reporting/viz/yolo_train_val
       hash: md5
-      md5: 7b3748c47c77eb11844bf0628927c3b4.dir
-      size: 3658088575
+      md5: 922d0a257ec915d91bb299f11df8ae27.dir
+      size: 3658189215
       nfiles: 16817
   visualize_yolo_test:
     cmd: uv run python scripts/visualize_dataset.py --dir-dataset data/processed/yolo_test
@@ -710,13 +714,13 @@ stages:
     deps:
     - path: data/interim/fp_sequence_embeddings/train
       hash: md5
-      md5: 51edec8cc1a3f1c8b93e7af595f80717.dir
-      size: 20216589
+      md5: d7c1ce9531de9fc9d89bf6f4305ba12d.dir
+      size: 20043098
       nfiles: 2
     - path: data/interim/fp_sequence_embeddings/val
       hash: md5
-      md5: aac0d5327b7b6ea3e05a66dccf154749.dir
-      size: 2491101
+      md5: 35a4487a1ba6a41ca2b529c6a699e6c3.dir
+      size: 2463314
       nfiles: 2
     - path: data/raw/fp/data
       hash: md5
@@ -742,16 +746,16 @@ stages:
       size: 348611
     - path: scripts/build_sequential_dataset.py
       hash: md5
-      md5: 7cb5dd1ef282a34d5cd0436a020f664f
-      size: 12785
+      md5: a279800d6a6e0a1c507b77256174d805
+      size: 11749
     - path: src/pyro_dataset/fp/lockfile.py
       hash: md5
       md5: cfd051c856b57783e3201555d1b4b8c6
       size: 2623
     - path: src/pyro_dataset/fp/selection.py
       hash: md5
-      md5: e9d2065653a9aafdc035c700fa2d678f
-      size: 7999
+      md5: e8b6b6673a571baeba4f94e18faeb98a
+      size: 10121
     outs:
     - path: data/processed/sequential_test
       hash: md5
@@ -760,23 +764,23 @@ stages:
       nfiles: 15320
     - path: data/processed/sequential_train_val
       hash: md5
-      md5: 9faac139ef157197191b84bc24bd1598.dir
-      size: 8555476683
-      nfiles: 144452
+      md5: ef44e3f9519456b587f424e27b5ad8a2.dir
+      size: 8556786817
+      nfiles: 144472
   build_toy_dataset:
     cmd: uv run python scripts/build_toy_dataset.py --yolo-src data/processed/yolo_train_val
       --seq-src data/processed/sequential_train_val --yolo-output data/processed/yolo_toy
       --seq-output data/processed/sequential_toy --ratio 0.05 --random-seed 0
     deps:
     - path: data/processed/sequential_train_val
       hash: md5
-      md5: 9faac139ef157197191b84bc24bd1598.dir
-      size: 8555476683
-      nfiles: 144452
+      md5: ef44e3f9519456b587f424e27b5ad8a2.dir
+      size: 8556786817
+      nfiles: 144472
     - path: data/processed/yolo_train_val
       hash: md5
-      md5: 038270d9336432fa9001cf92069731b2.dir
-      size: 2027739474
+      md5: a43f8eae91dd781aa91984d92bbe65d2.dir
+      size: 2028094915
       nfiles: 33635
     - path: scripts/build_toy_dataset.py
       hash: md5
@@ -790,8 +794,8 @@ stages:
       nfiles: 7728
     - path: data/processed/yolo_toy
       hash: md5
-      md5: 9bab5ff7b4ca24749be1120daa85f88a.dir
-      size: 101768902
+      md5: 257cdce256f23272ee72010f90fca5cd.dir
+      size: 101552974
       nfiles: 1683
   test_data_leakage:
     cmd: uv run pytest tests/test_data_leakage.py -v --junit-xml=data/reporting/leakage_results.xml
@@ -803,9 +807,9 @@ stages:
       nfiles: 15320
     - path: data/processed/sequential_train_val
       hash: md5
-      md5: 9faac139ef157197191b84bc24bd1598.dir
-      size: 8555476683
-      nfiles: 144452
+      md5: ef44e3f9519456b587f424e27b5ad8a2.dir
+      size: 8556786817
+      nfiles: 144472
     - path: data/processed/wildfire_yolo
       hash: md5
       md5: 7e25bea7a3afa9a9bf704b8e42c56a97.dir
@@ -818,8 +822,8 @@ stages:
       nfiles: 5281
     - path: data/processed/yolo_train_val
       hash: md5
-      md5: 038270d9336432fa9001cf92069731b2.dir
-      size: 2027739474
+      md5: a43f8eae91dd781aa91984d92bbe65d2.dir
+      size: 2028094915
       nfiles: 33635
     - path: data/raw/fp/registry.json
       hash: md5
@@ -857,12 +861,12 @@ stages:
       size: 1318880
     - path: scripts/embed_fp_for_selection.py
       hash: md5
-      md5: 27c76f4126615f2be3fb4b798e8c55ef
-      size: 8024
+      md5: ba8c3fb00f8f15b129aa481889d816b8
+      size: 8525
     - path: src/pyro_dataset/fp/selection.py
       hash: md5
-      md5: e9d2065653a9aafdc035c700fa2d678f
-      size: 7999
+      md5: e8b6b6673a571baeba4f94e18faeb98a
+      size: 10121
     outs:
     - path: data/interim/fp_sequence_embeddings/test
       hash: md5
@@ -871,13 +875,13 @@ stages:
       nfiles: 2
     - path: data/interim/fp_sequence_embeddings/train
       hash: md5
-      md5: 51edec8cc1a3f1c8b93e7af595f80717.dir
-      size: 20216589
+      md5: d7c1ce9531de9fc9d89bf6f4305ba12d.dir
+      size: 20043098
       nfiles: 2
     - path: data/interim/fp_sequence_embeddings/val
       hash: md5
-      md5: aac0d5327b7b6ea3e05a66dccf154749.dir
-      size: 2491101
+      md5: 35a4487a1ba6a41ca2b529c6a699e6c3.dir
+      size: 2463314
       nfiles: 2
   materialise_annotator_sequences:
     cmd: uv run python scripts/materialise_annotator_sequences.py --export-dir data/raw/pyro-annotator/export

dvc.yaml

@@ -39,13 +39,15 @@ stages:
       --data-dir data/raw/fp/data
       --wf-dataset data/processed/wildfire_yolo
       --embeddings-dir data/interim/fp_sequence_embeddings
+      --recurring-objects data/raw/pyro-annotator/recurring_objects.json
       --output data/processed/fp_yolo
       --random-seed 0
     deps:
       - scripts/build_fp_yolo_dataset.py
       - src/pyro_dataset/fp/selection.py
       - data/raw/fp/registry.json
       - data/raw/fp/data
+      - data/raw/pyro-annotator/recurring_objects.json
       - data/processed/wildfire_yolo
       - data/interim/fp_sequence_embeddings/val
       - data/interim/fp_sequence_embeddings/train