Decide a replay is too short before registering its match

[bill-rich]

What

upload_replay registered the match, computed its composition and invalidated the caches, and only then asked match_from_replay whether the replay was long enough to count, raising 422 when it wasn't:

_, created = replay_manager.register_match(db_match)
replay_manager.compute_and_save_composition(db_match.match_id)
invalidate_match_caches()

match_info = matches.match_from_replay(result.replay)
if match_info is None:
    raise HTTPException(status_code=422, detail="Replay is too short or invalid")

The writes had already happened by the time we decided the upload was invalid, so a short game landed half-present: reachable through /api/details and /api/replays, absent from /api/matches/by_date (get_match_infos lists with a 2.0 minute floor), and reported to the uploading client as a flat failure.

Match 866608453 from 2026-08-21 is a live example: a 42 second game someone quit out of. cncstats holds stats for it, radarvan answers /api/details/866608453, and it shows up in no date listing. It also makes "did every match land?" reconciliation between the two servers lie.

Change

Run the validation before any write, and register nothing when it fails. The parsed .rep and .json are still persisted by upload_and_parse, so the raw artifact survives either way; only the match row becomes conditional, which is what every read path already assumes.

Testing

New tests/test_upload_replay.py pins the ordering: a short replay must leave register_match / compute_and_save_composition uncalled, and a normal one must still register. Verified the first test fails against the old ordering and passes with the fix.

Full suite: 226 passed, 5 skipped. ruff check radarvan/ and mypy radarvan/ clean.