variants/linux: block on the LoRa IRQ in poll() instead of spinning; non-blocking hardware CAD

[mmmorks]

Staged on the fork for review. Final destination: l5yth/meshcore-linux linux. Base here is pr/03-sx1262-parity so the diff shows only this change.

Summary

Builds on l5yth#24. LinuxBoard::sleep() from l5yth#24 stays, and so does the delay(1) in the non-powersaving branch; this adds the piece that lets meshcored idle on the radio: the main loop blocks in poll() on a libgpiod edge-event descriptor for DIO1 (plus anything else it will drain that iteration) until the IRQ fires or a bounded timeout elapses. Idle CPU goes from a busy core to well under 1% with edge detection, and roughly 1–3% in the polling fallback, with no added packet latency — DIO1 carries both RX-done and TX-done, so packet events wake the loop immediately. On Linux the 1 ms delay(1) is now redundant (the poll() wait follows it); it is left in place because the other targets in this tree rely on it.

The same edge descriptor also fixes hardware CAD (set cad on): RadioLib's scanChannel() busy-spins on digitalRead(DIO1) with no deadline, which burned a core for every scan and would turn a GPIO read that starts failing into an unbreakable hang. performChannelScan() is overridden to start the scan, sleep on the edge with a deadline derived from the active SF/BW, and read the verdict over SPI regardless.

What changed

• src/MeshCore.h: virtual void MainBoard::idleUntilEvent(uint32_t max_wait_ms), default no-op, with the implementer's contract written down (do not lose an IRQ already asserted on entry; do not wait on a descriptor the caller will not drain). ESP32Board/NRF52Board already implement the same idea inside their sleep(), gated on the powersaving pref; this is the unconditional, always-safe variant. Boards that do not implement it keep the historical busy loop.
• examples/simple_repeater/main.cpp: calls board.idleUntilEvent(IDLE_MAX_WAIT_MS) (50 ms, overridable) at the end of loop(), after the existing powersaving block. The comment records why 50 ms: it is the floor of the delayed-inbound queue, and everything faster arrives on the IRQ.
• variants/linux/LinuxEventLoop.{h,cpp}: pure-POSIX poll() wrapper over an abstract LinuxEventSource, so it compiles and is unit-tested on a host with no libgpiod. wait() sleeps 1 ms and reports 0 on POLLNVAL/POLLHUP/non-EINTR failure, because poll() returns a positive count for those and a single closed descriptor would otherwise reinstate the busy loop.
• variants/linux/EventGPIOPin.{h,cpp}: an ardulinux GPIOPin that is also a LinuxEventSource, exposing a pollable rising-edge descriptor. Supports libgpiod v1 and v2 (selected by preprocessor check). setPinMode() re-applies edge detection, because RadioLib's pinMode(irq, INPUT) in SX126x::begin() would otherwise drop it. If edge detection is unavailable it degrades to a plain pin and the loop falls back to a 1 ms poll timeout; startup logs which mode the IRQ pin ended up in.
• variants/linux/LinuxBoard.{h,cpp}: the IRQ pin is bound through EventGPIOPin; idleUntilEvent() and waitForRadioIrq(). Both re-read the IRQ level via digitalRead() before blocking — in ardulinux that refreshes the cached level gpioIdle() fires the ISR against, and it is what lets the ceiling be longer than a packet's airtime without silently stopping RX on a latched-high DIO1.
• variants/linux/LinuxRadioWait.{h,cpp}: waitForIrqAsserted() (depends only on the event loop and CLOCK_MONOTONIC, so it is native-testable) and cadTimeoutMillis() (8 symbol times, twice RadioLib's 4-symbol scan, plus 20 ms).
• src/helpers/radiolib/LinuxSX1262Wrapper.h: performChannelScan() override = startChannelScan() → wait on the edge with the deadline → getChannelScanResult(). Relies on CAD_DONE being in the DIO1 routing mask alongside CAD_DETECTED, so a free channel arrives as an edge rather than a timeout (checked against the pinned RadioLib). symbolMicros() moves to RadioLibWrapper so the CAD deadline and calcMaxPacketMillis() share one formula.
• Tests: test/test_linux_event_loop (anti-spin properties: stale-descriptor backoff, filtered POLLIN count, POLLHUP handling) and test/test_linux_radio_wait (deadline governs under EINTR, each wake re-reads the line, INT_MAX clamp). platformio.ini's native env adds -I variants/linux and the two sources.
• README: ## Operation with "Idle CPU usage" (what the startup and degradation log lines mean) and "Channel Activity Detection" (how it composes with int.thresh, and the scan-duty cost at high SF); a Known Gaps entry that the libgpiod v2 path is compile-verified only.

Why

ardulinux skips its loop sleep whenever real hardware is bound, so before l5yth#24 meshcored burned 100% of a core; l5yth#24 made that a 1 ms sleep per iteration, which is a fixed-rate poll of the radio rather than an idle. Blocking on the IRQ line is what an event-driven daemon should do, and it is also the only way to make CAD non-blocking without giving up the synchronous "is the channel clear right now" answer isChannelActive() has to give (there is nothing to overlap with anyway — CAD puts the modem in standby).

How it was tested

• Run on a Pi with the Waveshare SX1262 HAT (libgpiod 1.6.3, bookworm): idle well under 1% CPU with edge detection, roughly 1–3% in the polling fallback (the figures the README quotes). Expected loop-wakeup rate at the 50 ms ceiling is ~20 Hz, down from ~88 Hz with the previous 10 ms ceiling. The libgpiod v2 path (trixie) is compile-verified in the container only.
• Native suite passes with the two new test groups (real signals are driven through the wait to cover EINTR).
• Clean linux_repeater build for arm64 in the container.
• The CAD path is covered by the native tests and a clean build; its on-air behaviour has not been measured against a second node.

Dependencies

Stacked on pr/03-sx1262-parity (the CAD override lives in the reworked wrapper). Review the last commit only until that merges. pr/06-gps adds lines next to this PR's platformio.ini native-env lines; whichever lands second will need a trivial rebase.

Shared code touched

• src/MeshCore.h (new virtual with a no-op default)
• examples/simple_repeater/main.cpp (one call at the end of loop(); a no-op on every board that does not implement it)
• src/helpers/radiolib/RadioLibWrappers.{h,cpp} (symbolMicros() extracted from calcMaxPacketMillis(); no behaviour change)
• platformio.ini (native test env only)