ToO mode: ~40x faster Target runs (ephemeris grid, dedup, serial EAS below threshold)

[Areustle]

Summary

Target (ToO) mode was dominated by astropy coordinate work, not shower physics: at 1M thrown, 97% of a 28.9 s run was get_body/AltAz evaluated independently at every thrown time. Three commits, each A/B'd against the previous state:

1. sun_moon_cut fetches each ephemeris once (dd39545). It called get_body four times per evaluation (twice again inside moon_phase_angle) and built three identical AltAz frames. Bit-identical output; cut stage 1.67x faster.
2. Sky positions on a coarse time grid + cubic spline (1055862). New simulation.target.ephemeris_step (default 60 s; 0 = exact per-time evaluation, the old behaviour). Source AltAz is interpolated as a unit vector (azimuth wrap-safe); sun/moon altitude and phase angle directly. Scalar/few-point queries still evaluate exactly. At 200k thrown: source error 2e-8 rad (astropy's own noise floor — 300 s gives the same), sun/moon 2e-10 rad, zero mask flips for steps up to 900 s. End-to-end 1M thrown: same rows and passing events, OMCINT rel. diff 2.6e-11, 25.1 s → 1.7 s.
3. EAS optical runs in-process below a valid-shower threshold (dc591c2). The dask cluster's ~1 s spawn/teardown was half the remaining wall time. CphotAng.__call__/EAS.__call__ gain serial=True (same packed chunk kernel, whole batch, this process); compute() decides after geometry using new simulation.eas_parallel_threshold (default 150 000; 0 = always cluster). Bit-identical both ways in Diffuse and Target at all sizes tested; break-even measured at ~150k valid (≈11 µs/shower serial vs ≈1.3 s + 2.7 µs/shower cluster).

Net: 1M-throw Target run 28.9 s → 0.74 s; large Diffuse runs unchanged (still use the cluster).

A/B evidence

Serial vs cluster, seed-matched, every column + MC meta array_equal:

mode	valid	cluster	serial
Diffuse	99k	1.88 s	1.10 s
Diffuse	992k	3.99 s	11.36 s
Diffuse	3.97M	10.9 s	45.5 s
Target	39k	2.27 s	0.74 s
Target	194k	2.88 s	3.07 s

Also prototyped (not adopted) a closed-form ICRS→CIRS-once + Earth-rotation-angle transform for the source: 2.5e-6 rad error (polar motion + diurnal aberration missing), and with ERFA apio13/atioq 9e-7 rad and slower than the spline. The grid already exploits the same slow/fast structure and reaches astropy's noise floor.

Reviewer notes

• Two new config fields, both defaulting to the fast path with an escape hatch to old behaviour (ephemeris_step = 0, eas_parallel_threshold = 0). test_config serialization expectations updated accordingly.
• Public signatures of get_sun, get_moon, moon_phase_angle(time) are unchanged; new sun_moon_state(time) and phase_angle_from_bodies(sun, moon) helpers.
• New test test_ephemeris_grid_matches_exact pins interpolated vs exact positions (<1e-6 rad) and identical cut masks on 5000 random times.
• Ruff flags a pre-existing unused refractive_index import in cphotang.py; left alone as unrelated.

Test plan

• pytest test/core test/simulation/geometry test/simulation/eas_optical — 162 passed
• black / flake8 clean on changed files
• A/B scripts above (scratch, not committed)

🤖 Generated with Claude Code