v0.1.0

Rolls up 0.2.0 (never released separately) and 0.3.0.

Highlights

• Automatic pipeline: dopplergeo detect|locate and the Hermes tools doppler_detect / doppler_locate. No hand-picked bands. Fetches adsb.lol ADS-B (ODbL), Copernicus DEM and Open-Meteo, with a cache (DOPPLER_GEO_CACHE) and --offline.
• Track-based scoring: emission-time-correct curves from real ADS-B tracks, a joint likelihood with a silence penalty, a coarse-to-fine posterior with 68 / 95 % areas, and a confidence: ok|low flag. The areas are calibrated with temper_extra=4.
• Broadband (experimental, --broadband / broadband: true): detects jet passes from the ground-reflection (Lloyd's mirror) multipath delay. Cepstrogram, Viterbi track, straight-pass fit, after Lo & Ferguson (IEEE TAES 2003). See docs/broadband.md.
• New CLI flags --broadband and --temper-extra, plus matching tool parameters.
• Eval harness scripts/eval_batch.py and docs/evaluation.md.

Honest results (real recordings with a known place and time; ±30 s; 15 km search)

• Tune, broadband on: Manchester field recordings went from 10.5 km to 0.19 km and from no events to 0.33 km. vny_0513 went from 15.4 to 2.95 km.
• Held-out (18 clips, one frozen run): 2 of 18 under 3 km (dfw 1.96, lgb 2.38), against 1 of 14 tonal-only. 11 solved against 8. The median of solved runs is 13.2 km (12.1 before). ads_0821b (7.9 → 17.2 km) and syd (11.0 → 18.6 km) regressed. Of the four ok results, lgb (2.4 km) and dub (4.7 km) are roughly right, and osh (13.2 km) and muc_ap82011 (13.1 km) are wrong. The padded 95 % area holds the truth in 8 of 11. The Edinburgh, The Hague and Munich-north clips and most livestreams give no events.
• The method works on clean multi-pass clips but does not yet generalise. Treat any output as a hypothesis to confirm.

Notes

• Network use is disclosed in SECURITY.md. Use only for forensics, research or your own recordings.
• Full changelog: CHANGELOG.md