feat(observers): per-observer activity endpoint

[MrAlders0n]

Summary

GET /api/v1/observers/{observerId}/activity?range=24h&interval=15m — a per-observer, time-bucketed view of what an observer heard: observation counts, LoRa airtime, SNR/RSSI aggregates and a whole-range payload-type breakdown. Feeds four new charts on the Stats → Observer tab.

Design

• Airtime is computed at ingest and stored on packet_observations.airtime_ms (RadioLib getTimeOnAir, MeshCore preamble rule, internal/lora). Ingest has the raw frame so the length is exact. Rows without radio params stay NULL.
• No request path joins back to packets — the same join that overran /dev/shm for the payload breakdown before migration 015.
• Intervals ≥ 1h read an hourly matview mv_observer_activity_hourly (observer, payload type, hour) over 30 days, refreshed with the other views. 5m/15m read packet_observations over idx_observations_observer, capped at a 48h window.
• Cached 60 s per observer/range/interval.

Migrations

• 030 adds the column (ALTER only, so the lock is released immediately).
• 031 backfills the last 7 days with a throwaway SQL copy of the formula, like 015.
• 032 creates the matview + unique index.

Verify before merging

These have only been reviewed, not executed — there is no Postgres on the dev box:

1. Apply 030–032 on a prod-sized copy; time the backfill UPDATE and the matview build, and time one REFRESH … CONCURRENTLY (this view is keyed by observer, so it is larger than the IATA-keyed ones).
2. EXPLAIN ANALYZE the raw 24h/15m and hourly 720h/6h queries on a busy observer; both should use idx_observations_observer and never touch packets.
3. Spot-check a few backfilled airtime_ms values against lora.TimeOnAirMs.

Tests

Formula spot checks (SF7/62.5/CR5/50 B → 244.224 ms; SF10 trips LDRO), end-to-end frame-length reconstruction for flood, transport and 2-byte-hash paths, gomock store tests for both paths and all null gating, handler tests for every validation rule. Swagger regenerated. Docs in beacon-docs api_contract.md.