Thread the altitude layer set through the associator

[jehanazad]

Why

An n=2 solve is exactly determined in (x, y) once altitude is pinned, so its position error is its altitude error. On a 20-minute capture against ground truth (dark lane), n=2 solves had a median altitude error of 1.46 km (p90 5.25) and a position error of 1.61 km when the altitude was within 1 km of truth vs 2.71 km when it was not (n=3, for contrast: 0.36 vs 2.54 km).

That altitude comes from this library: compute_overlap_zone precomputes each node pair's overlap on a horizontal lattice at a fixed set of altitude layers, and detection_association picks the best grid point plus a delay-residual-weighted mean altitude across them. The horizontal step is not the limiter — the LM solve converges from a 3 km start. The altitude ladder is, and it was hardcoded at six layers whose widest gap (1.5 → 3.0 → 5.0 km) is 2 km.

What

• altitudes_km is now a parameter of InterNodeAssociator, threaded into compute_overlap_zone on both zone-build paths (register_node and the rebuild), exactly the way grid_step_km already is.
• DEFAULT_ALTITUDES_KM keeps the historic (1.5, 3.0, 5.0, 7.0, 9.0, 11.0) as the library default, so every existing caller, unit test and the offline bench see the identical grid they saw before. The deployment overrides it.
• Comments that named the specific six layers (5, 7, 9, 11, "six altitudes") now describe the configurable ladder instead of a frozen one.

Cost

Linear, and nothing else moves: the columns are altitude-independent (the property the both-beams restructure rests on), so a twelve-layer ladder emits exactly 2× the grid points of a six-layer one over the same geometry and pays 2× the precompute. Measured on the test fixture pair: at grid_step_km=5.0, 66 → 132 grid points (4.2 → 4.7 ms); at the production grid_step_km=3.0, 180 → 360 grid points (9.3 → 13.3 ms). The per-zone build is dominated by the once-per-column both-beams test, so the wall cost grows well under 2×.

Tests

108 passed across test_association.py, test_detection_association.py, test_coverage_prior.py, test_bistatic_footprint.py. Two new ones pin the 2× ratio and that the associator's configured ladder actually reaches the zones register_node builds — a layer set that took effect on only one of the two build paths would be worse than none. ruff check / ruff format --check clean.

Consumed by the retina-server PR "Finer altitude layers and altitude continuity for n=2 solves".

🤖 Generated with Claude Code