Evidence, workaround and suggested fix for each: pilot-link's macOS USB
open race and dead debug logging, PilRC's corrupt NFNT header, the
prc-tools multi-section traps, MeshCore's radio params not retuning until
reboot, the clock persisting through contact lastmods, and stale bytes in
unused GET_CHANNEL replies. Nothing has been filed yet.
Code was 28.8 KB of the ~32 KB a single 68k code resource can reach.
visormesh.def declares section vmcode2; multigen generates the stub and
linker script; build-prc takes the name, creator and backup bit from the
.def. protocol.c and ui/text.c move there (segment 1 now 24.3 KB, segment
2 4.9 KB). Every function in them, static ones too, is declared VM_SECT2
so calls from segment 1 go through the __text__vmcode2 jump base, and
their two static const tables became plain data, since const data stays in
segment 1 where PC-relative code in segment 2 can't reach it. The rules
are in src/sections.h. Checked in the objects: no .text left in the moved
files, cross-segment calls go through the section bases.
Palm OS 3.5's clock is local time with no zone, and SessionNow used it as
UTC: outgoing messages carried times hours off, contact ages all read
'now', and clock sync set the radio (which only moves forward) to the
wrong time. SessionNow now applies a UTC offset kept as an app preference
(default from prefMinutesWestOfGMT, editable in Settings as hours, quarter
hours allowed); display converts back with SessionLocalPalm. Changing the
offset re-syncs the radio clock.
Contact, channel and node names are converted from UTF-8 when stored;
names with emoji showed missing-glyph boxes.
The transcript wrapped after a few characters on the Visor: gSmallFontWidths
was const, so prc-tools placed it in .text, but text.c reached it through an
extern declaration A5-relative (lea %a5@(0)) as if it were data, and read
garbage widths. The same applied to gPresets. Both generators now emit
non-const tables, with a note why. Host tests can't see this.
Home is now the conversation list (channels in slot order, then direct
conversations most recent first) with markers, unread badges, failed
marks, ages and previews; it connects on launch with the saved link.
The conversation screen shows day dividers, sender chips, time column,
wrapped text and SNR or delivery marks (pending, waiting for ack, acked,
failed), with a compose field, a UTF-8 byte counter, Send (or Graffiti
return), and tap-to-retry on failed messages. Contacts pins our node and
shows type marks, heard ages and routes; Message opens a DM. The old main
screen becomes Node with Advert and Done. One Options menu everywhere;
while connected the event loop polls every half second and posts a form
update when messages or contacts change.
Reception stayed dead after switching to PhillyMesh until the XIAO was
rebooted: SET_RADIO_PARAMS saves and reports the new values, but the
receiver only retunes on a restart. After a radio change, Settings asks
to reboot the radio (CMD_REBOOT + "reboot"; no reply), waits for it and
reconnects.
Query radio on the Visor Neo got DEVICE_INFO from the XIAO via USB
cradle -> visor_bridge.py -> FTDI -> D6/D7 UART. The app's HsUd driver
sends on the Visor's HOTSYNC USB port (2), not GENERIC, so the bridge
listens on every port and replies on the one in use; stalled endpoints
are cleared instead of ending the session. The app waits after SrmOpen
and retries sends that time out while the bridge attaches. make install
stops the bridge, which would otherwise grab the Visor during HotSync.
APP_START/SELF_INFO, GET_CONTACTS/CONTACT/END_OF_CONTACTS and
SYNC_NEXT_MESSAGE with v3 contact and channel message frames, laid out
per MeshCore MyMesh.cpp. Little-endian fields are assembled byte by byte
so they decode correctly on the big-endian 68k. Builders are checked
byte-for-byte against frames the real radio accepted; SELF_INFO and the
empty contact list are real captures (key and name redacted), contact
and message frames are synthetic. companion_query.py gains --text.