Publish a synthetic node's declared cone as its detection area

[jehanazad]

Summary

Adds EmpiricalCoverageState.declared_wedge_polygon() and a per-node declared_geometry_is_truth flag on NodeAnalyticsManager.register_node, so a node whose declared cone is its detection area (a simulator node) is published as that cone instead of as its accumulated evidence. Real nodes are unchanged: evidence-only polygon, now tagged polygon_source: "evidence" ("learned" under FOV shadow/active).

Stacked on #28 (base branch is its head); merge #28 first and GitHub retargets this one to main.

Why

The simulator only emits a detection for an aircraft inside a node's declared cone (retina_simulation/world.py::_aircraft_in_detection_cone), so for a synthetic node the wedge is ground truth. The calibration points that feed the evidence polygon come from ADS-B hexes bound to tracks, and a steady share of those binds are the wrong aircraft. Measured on the test deployment 2026-09-13 from the persisted bins: one 42° node held 3,037 points of which 47 % lay outside its wedge (34 % ignoring the two edge bins); 55 out-of-wedge bins had reached FOV_OPEN_MIN_POINTS, and the published polygon covered 71 of 72 bearings. Across an eight-node sample 5–41 % of points were fully outside the wedge, yet even a 5 % node opened 15+ stray bearings, because three points open a bin.

Changes

• declared_wedge_polygon(step_deg=2.5): apex, both wedge edges exactly plus every step_deg between, apex — radius per bearing is _reach_at (the bistatic ellipse when TX + differential limit are declared, else the monostatic circle: the simulator's own range rule), no clamp multiplier, no bins read. Omni prior → closed ring, no apex.
• register_node(..., *, declared_geometry_is_truth=False): membership in _declared_truth asserted per registration, dropped on retire / reset / loss of geometry; a membership change invalidates the summaries cache.
• get_node_summary: declared nodes publish the wedge with polygon_source: "declared" regardless of n_points, and skip the evidence-derived FOV diagnostics; everyone else gains polygon_source.

Test coverage

• tests/test_declared_wedge.py: 26 new tests — vertex bearings within half-width and ranges at _reach_at (±1 %), both edges present, closure, omni ring, elliptical vs circular reach, manager publication for flagged / unflagged nodes, zero-point publication, flag drop on re-register and retire, cache invalidation.
• Full suite: 510 passed. pre-commit run --all-files: ruff, ruff-format, vulture, ruff-config all pass.
• Independent cross-check (not in the repo): the wedge built from the real persisted bins of the 42° node above agrees with SimulationWorld._aircraft_in_detection_cone on 25,919 / 25,921 grid points; both disagreements sit on the boundary.

Review notes

• Vertices are placed with the flat offset_latlon step every polygon in the file uses, so reading a vertex back spherically shows a ~0.2° convergence term at 50 km; the tests measure spherically on purpose and carry a documented 0.25° slack.
• The flag is not persisted: it describes who registered the node and is re-asserted on every registration; detection areas are not persisted either, and a summary needs one before any polygon is published.
• The same wrong-aircraft bind rate inflates real nodes' evidence polygons; that is a calibration-attribution follow-up, deliberately not touched here.

🤖 Generated with Claude Code