Four root causes behind zero moments, measured against your production data

[ralyodio]

You said stop working around it. So I pulled the real project database off the prod volume and ran the actual scorer over your 9961 tracks. Every number here is from Sam's game, not a fixture.

Four root causes

1. Tracks accumulate. Detection appended instead of replacing, so each re-analysis layered another copy over the last.

9961 stored tracks  ->  2790 distinct
top of the list: trk_850a27bc / trk_8a58250b / trk_e67f99f7
                 all 33.334s, all 364 points — the same track, three times

clearTracks now replaces a video's tracks. Because focal_track_id is a bare column with no foreign key, it also clears the binding and warns you to re-identify — otherwise scoring keeps an anchor pointed at a deleted row.

2. Only the first ball track was ever read. tracks.find(t => t.className === 'ball'). Your detector emits the ball as 56 fragments totalling 10.1 seconds of 300, longest 0.4s — so the signals carrying 0.45 of the weight read one arbitrary fragment. All fragments now count; when several are live, the one nearest the athlete wins.

3. An unseen ball scored zero instead of nothing. That kept 0.45 of weight in the denominator contributing nothing to the numerator, for 97% of the game. Absence of evidence isn't evidence of a dull moment. Worth 0.251 → 0.305 on your data.

4. Sam is not one track. The tracker splits a child into fragments; the picker showed him as several people; binding one followed 23.9 seconds of a 300-second game. You spotted this — "there's 2-3 photos of him".

An athlete can now be identified across several fragments. tracks.athlete_id already existed and nothing wrote it. Scoring stitches them into one series, keeps the more confident observation where fragments overlap, and refuses to interpolate across a gap — a straight line through a hole puts a child on the court while he sat on the bench. The virtual cameraman still interpolates freely; the cap is focal-only.

Measured on your game

Sam = 1 track (as prod is today)   best 0.305   0 moments
Sam stitched from 3 tracks         best 0.419   3 moments
Sam stitched from 12 tracks        best 0.435   5 moments
                                   20s-27s 0.435 [player_ball_proximity,
                                                  player_acceleration, high_motion]

Coverage 23.9s → 93s → 99s. Caveat worth stating: I selected those extra fragments by duration, not by verifying they are Sam. The mechanism is proven; the exact scores depend on which crops are actually him.

What this does not fix

The detector cannot see a basketball. 1920×1080 → 540p proxy → the model's fixed 416×416 input makes the ball roughly six pixels, below YOLOX-Tiny's finest stride of eight. It is under the model's resolution floor, which is why 10 seconds of 300.

That caps this footage at 0.435 against a 0.35 threshold, on player acceleration and scene motion alone — moments clear by 0.07. Thin. The durable fix is detection at higher effective resolution: a 640-input model, or tiling each frame into quarters at 416 (≈4× inference, ~8 min for a 5-minute clip). I did not guess at that here because it needs validating against your footage, not asserted.

Also unused: your video has an audio stream and nothing extracts it, so audio_spike (weight 0.1) has been permanently dark.

Verification

423 tests pass (14 new), lint and typecheck clean. The prod numbers come from explainScoring and computeMoments run over the exported tracks.

One correction made along the way: my first version capped interpolation globally and broke the virtual cameraman, which legitimately bridges sparse keyframes. The cap is now opt-in and applies only to the focal athlete.