Split contaminated position clusters instead of picking the loud track

[jehanazad]

What

format_track_pairs_for_solver merged every pairing within 6 km into one solver input, and where a node turned up in that cluster with two different tracks it kept the one with the higher SNR. A node's tracker gives one track per aircraft, so that case is two aircraft — and resolving it by loudness hands one aircraft's candidate a measurement belonging to the other.

Worse, the union-find that builds the cluster is transitive, so "within 6 km" chains. On the 15-node bench scene the baseline emitted 208 solver inputs spanning 19–23 single-node tracks. An aircraft has one track per node, so a 23-track input at 10 nodes was describing at least three aircraft as one.

Four changes, each separately gated and separately measured on backend/scripts/association_bench.py --mode track --cv-fit deferred (production's shape: cv_fit=None, so the chi2 exclusivity in _pair_tracks stage 2 never runs live):

1. _partition_cluster replaces the SNR pick. Every connected group is partitioned into sub-clusters that are consistent by construction, and membership is tested against every pairing already in a sub-cluster rather than one of them: same node-to-track assignment, within merge_dist_km, implied-velocity consistent. That is what bounds each emitted input's diameter instead of letting it grow with the chain. All sub-clusters are emitted — this stage cannot tell which aircraft is real, and suppressing the runners-up is how the SNR pick went wrong. New counter cluster_splits.
2. _drop_velocity_conflicts: pair-level exclusivity on the Doppler-implied velocity, gated on cv_fit is None. Not a standalone test — the implied speed measured 0% power against real cross pairings — but two pairings that share a track and imply velocities that cannot both be true are competing claims, and the smaller coarse delay residual wins. It abstains wherever either side has no inference, so it cannot cost the recall the earlier delay-residual assignment cost. Losers counted as track_pairs_superseded.
3. Merge criterion _MERGE_DIST_KM 6.0 to 3.0 — one association grid step, not two — and a union edge now also requires implied-velocity agreement. Swept 6.0 / 4.5 / 3.0 on the bench; monotone across the whole range with the real-track count identical at every point, so there is no trade to make.
4. chi2_per_dof stays None when nothing in the cluster was fitted — which on the deferred path is always.

Why

Read-only audit on the live test droplet, 156 GT-matched dark records:

• 58–65% of dark solver candidates carried a node that could not see the aircraft; 83/129 of those foreign nodes were in the cone of another aircraft.
• 31 of the 34 candidates with pre-trim rms > 3 µs were contaminated.
• The solver trimmed 547/1136 successes, and 33 of 69 trimmed nodes were legitimate.
• 45% of published dark solves still carried a foreign node; position error p90 4–5 km.
• Candidate node count was about equal to in-cone node count (5.7 vs 5.9) — contamination substituted wrong nodes for right ones rather than adding them.

Measured

15-node ring scene, 6 seeds, --mode track --cv-fit deferred. Full per-seed tables are in the retina-server PR that pins this commit.

	baseline	partition only, 6 km, no velocity	+ pair-velocity, 6 km	+ merge-velocity, 6 km	merge 4.5 km	shipped, 3.0 km
published contamination, per-seed mean	33.3%	32.7%	32.0%	30.2%	28.2%	25.1%
foreign nodes per published solve	0.80	0.80	0.76	0.70	0.63	0.52
ghost by track (pooled)	54.9%	56.8%	55.6%	54.9%	54.3%	50.0%
ghost by solve (mean of seeds)	5.95%	5.90%	5.93%	4.58%	4.23%	2.85%
real tracks per seed	3 4 6 6 6 7	3 4 6 6 6 7	3 4 6 6 6 7	3 4 6 6 6 7	3 4 6 6 6 7	3 4 6 6 6 7
widest solver input	23 tracks	10	10	10	10	10

Ghost rate by track and by solve is better or equal on every seed; the real-track count is identical on every seed at every leg; median matched error is flat (mean over seeds 1.61 → 1.62 km). The partition alone (column 2) is not a win — it only becomes one combined with the velocity work and the smaller radius.

On the 50-node scatter (droplet-like) scene: published contamination 12.7% → 7.6% pooled, foreign nodes per published solve 0.14 → 0.08, real tracks identical on every seed; two seeds regress on track-level ghost rate (0% → 17%, 11% → 20%, i.e. 1 and 2 extra false tracks).

The acceptance target (candidate contamination at or below 15%) is not met — it moves 31.9% → 28.7% on the ring scene and 20.3% → 20.0% on the scatter scene. The retina-server PR explains why: that metric's denominator is the number of candidates association emits, and breaking one 23-track cluster into its constituent targets plus the false pairings it was hiding emits several inputs where there was one, each of the latter scoring as 100% contaminated.

Verification

libs/retina-analytics suite: 425 passed. New tests: TestClusterPartition, TestVelocityExclusivity, TestVelocityConflict in tests/test_track_association.py.

Intentionally changed behaviour

TestFormatTrackPairsForSolver::test_cluster_reports_its_worst_fit used two pairings a1/b1 and a2/b2 at the same position to make one cluster. That is now (correctly) two targets, so the worst-fit rule would never be reached; the fixture becomes a1/b1 and a1/c1 — the same aircraft on three nodes, which is the case the rule is actually about.

🤖 Generated with Claude Code