The planner inlined the view CTEs into consumers and re-ran the links
aggregation once per candidate row (loops=13070, 24s+ for getViableLinks).
Wrapping each view body in WITH v AS MATERIALIZED forces one evaluation
per query: viable-links query 24.3s -> 3.1s, WS initial state <1s.
Also drops the dead terrain_profile_json column from node_identity_links
(no readers; frontend terrain comes from DEM tiles).
- channelRegistry: 39 validated channel secrets committed as the service
default (Public + 38 community/hashtag channels recovered via the
derivation audit); MESHCORE_CHANNEL_SECRETS env appends on top
- shared buildSummary/identifyChannel (single implementation for ingest
and offline tools)
- tools/backfillDecrypt: keyset-paginated retroactive decryption of the
stored raw_hex corpus (idempotent, resumable)
- feed: channel scope sidebar now lists every decrypted channel
Backend: schedule ONE final multi-observer resolution per path-bearing
packet after the propagation window closes (default 60s, calibrated on
real first->last observer spread p50 11s / p95 34s / p99 50s; median 17
observers per packet). The final runs with the COMPLETE observer set and
pushes into the resolve cache after invalidation, so the next API read
serves the full-propagation answer. mode=slow on /path-beta/resolve-multi
returns 202 {status:pending, remainingMs} while inside the window; new
/path-beta/slow-mode status endpoint; observability table
path_slow_resolutions (migration 034). Env knobs:
PATH_SLOW_MODE_ENABLED / PATH_SLOW_MODE_WINDOW_MS / PATH_SLOW_MODE_MIN_HOPS
/ PATH_SLOW_MODE_PENDING_MAX. Best-effort: restart loss is harmless.
Frontend: pinned packets resolve in slow mode (bounded 202 retry), local
path renders meanwhile, upgraded when the final lands. Live feed unchanged.
Tests: 4 new slowMode unit tests (dedupe, disable, bound, window timing).
Backend build clean, 257/258 (1 pre-existing workerPool timing flake on
base too); frontend build clean, 74/74.
gadgethd