Overhaul ADS-B pipeline: dead reckoning, signal-gap handling, stale-track rendering

[d3mocide]

What

End-to-end overhaul of the ADS-B pipeline (local SDR BEAST feed + OpenSky supplement) targeting the three reported pain points: dead reckoning, odd signal gaps, and frontend rendering of stale tracks.

Server-side dead reckoning (poller/normalizers/beast_decoder.py)

• Stale positions are now projected forward along the last known velocity via project_position() (which existed in beast_math.py but was never called). DF17 TC19 velocity messages usually keep decoding through CPR position gaps, so the projected track stays close to truth.
• Bounded by new ADSB_DEAD_RECKON_MAX_SECONDS (default 60 s); the stale threshold is now configurable via ADSB_POSITION_STALE_SECONDS (default 10 s). Altitude projects with the vertical rate.
• DR positions are flagged position_dr (plus position_age_s) and are estimates only: they never enter the trail ring buffer, DB observation history, or geofence entry/exit checks (guarded both at the publish call and inside db.write_entity_observation).

Signal-gap fixes

• Cross-source CPR seeding: OpenSky and ultrafeeder positions seed the decoder's Tier-2 local CPR reference (seed_reference()), so a single odd/even frame resolves a position the moment an aircraft (re-)enters SDR range — previously a cold lock waited up to ~60 s for a fresh even+odd pair. Local fixes stay authoritative; seeds apply only when the decoder's own fix is missing or stale.
• The registry tick loop now refreshes decoder entities on every snapshot tick, not only when new BEAST frames arrive — a total feed outage no longer freezes the published snapshot while DR is active.
• The CPR heading-consistency guard is bounded to gaps < 30 s; across longer gaps an aircraft may legitimately have turned, and the guard could previously reject the first good fix after a gap indefinitely.
• normalize_tar1090 fixes: alt_baro: "ground" no longer leaks the string into altitude and now correctly sets on_ground status (readsb has no on_ground key); seen_pos is honored (drop > 60 s, flag stale > 10 s); non-numeric track/gs/baro_rate are nulled.
• Redis-hydrated aircraft (position but no fix timestamp) now report position_stale until a real fix arrives, matching the hydration docstring's intent.

Frontend rendering

• PVB keeps projecting server-dead-reckoned tracks between reports — previously stale tracks froze (projectedSpeed = 0), then snapped to the next fix. positionDr joins the report key so DR↔fix transitions re-blend smoothly.
• Stale/dead-reckoned local aircraft render dimmed (like OpenSky supplements) so estimated positions are visually distinct from live fixes.
• Entity/Track types now carry position_stale / position_dr / position_age_s.

Relationship to PR #115

The raw-byte DF/length fast path from #115 (Jules) is folded into this branch verbatim (with credit) since this overhaul rewrites the surrounding ingest() lines — merging this PR supersedes the beast_decoder.py half of #115. The .jules/bolt.md log addition in #115 is untouched here, so #115 can still be merged for that file (expect a trivial conflict in beast_decoder.py if it merges second).

Testing

• 13 new decoder tests: fast-path rejection before any pyModeS call, dead-reckoning bounds (fresh/stale/capped/no-velocity/hydrated, altitude projection), and seed-reference arbitration.
• 9 new tar1090 normalization tests (ground handling, seen_pos staleness, non-numeric fields).
• Full poller suite: 155 passed. tsc --noEmit: clean. docker compose config: valid.

🤖 Generated with Claude Code

https://claude.ai/code/session_01G6ZwvPTHAZ76MxF6swwryH

---

Generated by Claude Code