Fix macro loading and UART data loss on wired pendants

[f1adang]

Summary

On an M5Dial wired to FluidNC's UART, the Macros screen showed "No Macros"
even with macros present, and the link carried a continuous flood of error:3.
Tracing it turned up six separate defects, most of them on the shared path that
streams $File/SendJSON, $Files/ListGCode and $File/ShowSome, so file
listing and file preview were affected too, not just macros.

Verified on hardware (M5Dial, FNC_BAUD 1000000, UART transport). Each commit
builds m5dial, cyddial, m5dial_trace and the macos simulator.

The defects

1. The WiFi stack starved the UART reader. fnc_poll() reads a single byte
then calls poll_extra(), which drove wifi_poll() unconditionally whenever
USE_WIFI was compiled in. At 1 Mbaud a byte arrives every 10 µs, so a
UART-connected pendant entered the WiFi stack ~100k times a second. The RX ring
overflowed and bytes vanished mid-document — runs of text disappeared, and the
tail of a status report ended up spliced into a JSON chunk:

[json] len=82 d=4 | d": "btnZX",    "name": "btnZX<Idle|WPos:59.591,6.000,8.000|...

Nothing in wifi_poll() is on the UART data path, so in UART mode it now runs on
a 20 ms timer. Telnet and ESP-NOW keep the per-byte call, where the socket refill
genuinely is on the data path.

Two buffers on the same path were also undersized for 1 Mbaud: the UART RX ring
was 256 bytes (~2.5 ms of headroom, less than one frame render) and the drain
loop in loop() stopped after 64 bytes per tick against 100-byte chunks. Now
2048 and 1024 — the loop already exits the moment RX is empty, so the larger
bound costs nothing at idle.

2. vsend_linef() formatted into a static buffer. send_line() →
fnc_send_line() spins in while (_ackwait) { fnc_poll(); } before it reads
the string, and fnc_poll() dispatches reports whose handlers call
send_linef() again. The inner call overwrote the buffer while the outer one was
still about to transmit from it, so commands went out mangled:

{"files":[],"path":"/sd$G","error":"Bad path"}

$Files/ListGCode=/sd lost its terminator and $G landed on the end of it. This
is the source of most of the error:3 (Bad $ statement) traffic. Long-standing
bug; it became reachable as handler-driven traffic increased.

3. Config-item requests retried forever. service_config_requests() re-sent
configRequests.front() every 500 ms until FluidNC answered with a matching
$name=value. A setting the machine doesn't have — an unconfigured axis, a key
this build lacks — is answered with a bare error:3 that never matches
parse_dollar(), so the item never left the queue. detect_homing_info() queues
twelve of these on connect, and since only front() is ever sent, one
unanswerable item blocked every request behind it while flooding the link.
Retries are now capped, and an error arriving while a query is outstanding drops
that item immediately.

4. show_error() made its own guard dead code. It reset the JSON depth and
then called file_request_failed_advance(), whose first act is to check
json_in_progress() — which could no longer be true. That guard exists because
FluidNC interleaves messages: an error from an already-failed request can land
after the next request's document has begun, and advancing there both tears
down a healthy document and skips past it.

5. Echo-off followed the build, not the transport. The Ctrl-L was guarded
#ifndef USE_WIFI and commented "UART only", but USE_WIFI means a network
transport is available, not in use — so a pendant built with it and wired to
UART stopped sending echo-off. With echo on, an echoed $... line arriving
mid-document trips the json_reset_depth() in handle_other().

6. preferences.json macros were never parsed. Three things:

• PreferencesListener is installed part-way through the document, on the
  "result" key of the $File/SendJSON wrapper, so its _level is relative to
  where it took over. preferences.json puts settings/macros at _level 1
  there, not 2 — and the _level < 2 bail swallowed the macros key before the
  _level == 2 check ever saw it.
• Unlike the other two listeners, startObject() never cleared
  _name/_target/_filename, so an entry missing action inherited the
  previous entry's filename — already prefixed — and got prefixed again,
  surfacing as //localfs/. An entry missing type inherited the previous type
  and was added when it should have been skipped.
• Only one key spelling was accepted. WebUI versions differ: older exports use
  filename/target (what MacroListListener expects), newer ones
  action/type. Both are now accepted, ESP is treated as a localfs target
  alongside FS, and a leading slash is normalised before prefixing so neither
  convention yields /localfs//foo.g.

Additionally, $RI auto-reporting kept emitting <Idle|MPos:...> on the same
channel during a transfer, so any byte lost to a stall welded the two together.
It is now suspended for the duration of a transfer, resumed on a condition that
holds down every exit path (including the list and preview transfers, which have
no chain state of their own). And service_macro_chain() puts a deadline on an
in-flight request so the menu can no longer sit on "Reading Macros" forever.

Also included

m5dial_trace env. FNC_RX_TRACE already existed but printed nothing on the
M5Dial — dbg_print/dbg_write compile to no-ops unless DEBUG_TO_USB is also
defined, which the m5dial env doesn't set. Trace output is buffered and flushed
only when nothing is streaming: printing inline is self-defeating, since
dbg_print blocks waiting for USB buffer space and 50 ms of blocking at 1 Mbaud
is ~5 KB of arriving UART data — more than the RX ring holds, so the act of
tracing destroys the transfer being traced.

Command macros are now runnable from the green button. A cmd: macro has no
file behind it, so invoke() with no argument (which means "open the file
preview") left green and touch doing nothing at all — such macros were reachable
only from the dial. onGreenButtonPress() also returned early unless state was
Idle, and $Job/Resume exists precisely for when the machine is not idle.

Not included

• No change to the !python ./git-version.py invocation. It fails on macOS,
  which has only python3, but hardcoding python3 would break Windows.
  The portable fix is extra_scripts = pre:./git-version.py — the script writes
  src/version.cpp and prints nothing to stdout, so it contributes no build
  flags and is purely a pre-build hook — but that has to be added to every env,
  so it belongs in its own PR.
• MacrocfgListener has if (++_level = 2) and if (--_level = 1) —
  assignment, not comparison, so both branches always taken. Left alone
  deliberately: the broken behaviour looks load-bearing, and correcting it needs
  a real macrocfg.json to test against. Not on the path this PR fixes.
• Building without USE_WIFI is currently broken on main: SystemScene is
  inside #ifdef USE_WIFI, but BrightnessScene.cpp:32-33 and
  DisplaySettingsScene.cpp:35,39 call activate_scene(&systemScene)
  unconditionally. Fixing it means deciding where those screens navigate back to
  when there is no System scene, which is a UI call.

---

🤖 Generated with Claude Code

[f1adang]

Yay, more vibe contributions ahoy! 🤪

PR containing some fixes against current main I discovered while debugging some issues with Brushograph remote control

[f1adang]

If you'd rather have this in separate PRs per commit, let me know