Both Clyde North observers stopped answering over WiFi after days of
uptime while LoRa and MQTT kept working. `memory` showed heap_min ~1KB
and a largest internal block of 19KB: the HTTPS listener was alive but
no mbedTLS handshake (~40KB contiguous) could be allocated.
The trigger was the Aug 9 lru_purge change: with a 2-socket pool, every
browser page-load burst evicted a live session and forced a fresh TLS
handshake, and the repeated 40KB alloc/free cycles fragmented internal
RAM. Before that change the pool simply jammed, so nothing churned.
- Send `Connection: close` and trigger a session close on one-shot
responses (pages, favicon, redirect, login, 401s) so page-load bursts
release their sockets immediately. The authenticated /api/* polling
connection keeps keep-alive so it does not pay a handshake per poll.
- Gate WebPanelServer::start() on internal heap headroom (56KB free /
32KB largest), retrying every 15s instead of every loop tick.
- Self-heal: when the panel is idle and the largest internal block drops
below 24KB, stop and re-create the server to return its pools.
- Count the MQTT teardown path that deliberately abandons a client on a
heap-integrity failure.
- Expose the above as `heals:`/`deferred:` in `get web.status` and
`leaked:` in `get mqtt.status`.
- Fix the stats page Channel/Gateway cells showing `--`: the /api/stats
summary JSON never carried channel, gateway health, or the watchdog
count (only the `get wifi.status` string did).
- Add eastmesh-tools/web-heap-check.sh to read heap/service health over
the API and optionally stress the panel with browser-style bursts.
- Release notes 2026.8.3 for both observer tracks.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
