fix(platform-import): adapt to new pyro-api sequence/detection schema

[MateoLostanlen]

Summary

The platform API restructured its sequence/detection fields, which silently broke make import-platform — every sequence ended up with KeyError on the new field names, swallowed by a per-sequence try/except, leaving 0 records from N sequences with no diagnostic. This PR adapts the import script to the current schema and tightens a few related correctness issues that codex review surfaced.

Schema changes addressed

• Sequence: camera_azimuth (= pose.azimuth at creation), sequence_azimuth (smoke cone direction, ignored), pose_id, cone_angle. Old single azimuth gone.
• Detection: bbox (string "[(x,y,x,y,c), ...]") + others_bboxes (siblings of the same image). Old bboxes and azimuth gone.

What's in the patch

• utils.py:to_record — rewritten for the new schema. Internal record key sequence_azimuth → camera_azimuth (the value is the camera/pose direction; the platform's sequence_azimuth is the smoke direction we don't want). Bbox parsing uses ast.literal_eval with defensive handling for flat singular tuples and malformed values; merges bbox + others_bboxes so each image carries all detected boxes.
• sequence_fetching.py:process_single_sequence_detections — dropped the bare try/except that returned [] and hid the schema drift. Detections are deduped by bucket_key (the platform emits one row per bbox, so without dedupe a multi-box image would be imported N times with N-box predictions each). Fetches min(limit + 10, 100) so dedupe doesn't under-fill, and preserves 0 = no limit.
• client.py:list_cameras — passes include_non_trustable=true. Without this, sequences from non-trustable cameras of the same org couldn't be matched against the camera index (your sdis-40 has 5 cameras total, not 4).
• shared.py:transform_sequence_data — rounds float camera_azimuth to int and wraps with % 360 (annotation API stores Optional[int]; round(359.6) = 360 is out of range).
• import.py / batch_import_local_yolo.py — clone path uses camera_azimuth; CSV consumer falls back to legacy keys for older exports.
• .gitignore — ignore local .claude/ worktree state.

Codex review

Multiple rounds. Net result: codex went from blocker findings → P2 → P2 false-positive. Findings addressed in commits: defensive bbox parser, bucket-key dedupe with fetch buffer, azimuth modulo, missed clone-path key rename, detections_limit <= 0 short-circuit. Final P2 (claim that platform doesn't return camera_azimuth) is incorrect — verified directly against the API response.

Test plan

• make import-platform DATE_FROM=2026-04-15 DATE_END=2026-04-15 --max-sequences 2 (smoke) — 1 sequence + 30 detections + 1 annotation, end-to-end clean
• make import-platform DATE_FROM=2026-04-15 DATE_END=2026-04-17 MAX_SEQUENCES=0 (3-day on clean DB) — 48/48 sequences, 978/978 detections, 48/48 annotations
• Re-run on populated DB — duplicates skipped cleanly, no errors
• Annotation API state verified: sequences == annotations in ready_to_annotate (no orphans)
• Pull pipeline (make pull-sequences) end-to-end — pending local stack bring-up (postgres volume mismatch unrelated to this PR)

🤖 Generated with Claude Code