writeStream() drained the whole queue in one call - "send every packet we can".
A client asking for the full config gets the node database, then the file
manifest, then the packet backlog and the position replay, and none of that
returns to loop(). On a full node database (120 entries) the dump runs past
eight seconds, so on RP2350, where rp2040Loop() arms an 8s hardware watchdog
and is the only thing that calls watchdog_update(), the board resets in the
middle of the manifest. Reproducible on every connection; with a small node
database the dump finished under the timeout and nothing looked wrong.
Measured on a pico2_w5500_e22: last loop iteration at millis=27242, ServerAPI
kept logging until uptime 35s, reset at 35.2s = 27.242 + 8.0.
Take a slice instead. The PhoneAPI state machine is resumable, so writeStream()
stops after STREAM_WRITE_BUDGET_MSEC and reports whether anything is left;
runOncePart() then asks to be re-run immediately rather than sleeping out
readStream's idle delay, so the dump keeps its throughput while loop() gets to
feed the watchdog between slices. Backpressure on a retained frame still
returns the normal delay - re-running at once would just spin on a full
transport.
Verified on hardware: 10/10 full config dumps against a node with 120 entries,
no resets, dump still completes in ~8s.
Carlos Valdes
d89f3eb8fc