[Strycher]

Closes #290, closes #291.

Two firmware defects found on the UM FeatherS3D bench on 2026-09-08, both pre-existing on main and independent of the board, plus the display module's official pinout and the FeatherS3D power-path facts in the docs.

#290 — a tap wakes a sleeping TFT

loop() parks lv_timer_handler() while the TFT sleeps, and LVGL's indev callback was the only reader of the FT6336U, so once asleep only the three wing buttons could wake the panel. Hidden until now because TFT sleep defaults to "always on"; the owner enabled it.

• touch.cpp/touch.h: touchPollForWake() — the CTP_INT falling-edge ISR sets a flag (no pin read, nothing outside IRAM), consumed on the next loop pass; a 100 ms ctp.touched() poll is the fallback for a bench without CTP_INT wired. The waking tap is swallowed until the finger lifts (touchWakeSwallow, called from the LVGL read callback), capped at 2 s so a chip stuck reporting "touched" cannot disable input.
• src.ino: while tftSleeping, touchPollForWake() → wakeAllDisplays(), the same path the buttons use. Logs [SLEEP] Woken by touch.

Hardware evidence (um_feather): cf3f8fd (poll only) woke after 90 s, ~4 min and an overnight idle, but needed several taps; ea4a99b/e3818d3 (edge-driven) woke on the running image, logged. The post-review edge-mode change (FALLING instead of CHANGE+read) is built but not yet flashed.

#291 — GPS parser crash on a short RMC sentence

parseNMEA declared an uninitialised fields[], nmeaParse filled only nFields slots, and the RMC branch read fields[3..7] after a nFields >= 3 gate. A sentence cut short by a UART hiccup with status A read stale stack pointers and faulted at strlen(fields[7]): LoadProhibited, EXCVADDR 0, decoded with addr2line against a rebuild of the crashed commit (parseNMEA gps.cpp:138 ← readGPS ← loop). Three identical panics in one night, 1.5–3.5 h apart.

• gps.cpp/gps.h: every unfilled slot points at an empty string; fields[] is now const char* so a write through it cannot compile.

Hardware evidence: e3818d3 has run since 12:07 EDT with no GPS panic (the old image failed within 1.5–3.5 h every time).

Docs (#290 branch)

• docs/um-bench-wiring.md and docs/HARDWARE.md: the display module's official 14-pin order from the LCDWIKI MSP3526 manual (header J2 and the 14P 0.5 mm FPC P2 carry the same signals in the same order, verified from the schematic), CTP_RST → hole 14, the manual's 5 V recommendation, and the FeatherS3D power paths from its schematic (5V pin = VBUS via a diode, no boost on board).

Gemini adversarial review (gemini-2.5-pro, docs/llm-consultations/2026-09-08-touch-wake-gps-parse-*.log)

#	Finding	Action
1	digitalRead() inside the IRAM ISR flagged as non-IRAM	The claim is wrong for core 3.3.8 (__digitalRead is ARDUINO_ISR_ATTR), but it has a log_w path, so the read is gone anyway: the interrupt is now FALLING and the ISR only sets the flag. Fixed.
2	Non-atomic check-then-clear of the flag can lose a second edge	A lost second edge changes nothing: the first edge is already waking the panel. Justified, documented in code.
3	Swallow-until-release never clears if the chip reports touch forever	2 s cap on the swallow. Fixed.
4	(char*) cast off the const empty string is a latent write hazard	fields[] is const char* end to end; no cast. Fixed.
5	talker[2] read without a length check	isRMC is only true when strstr(talker, "RMC") found the substring, so talker has at least three characters on that path. Justified.

Builds

um_feathers3 and feather_s3 both green on the tip; the Adafruit env is unchanged in behaviour. pytest scripts/ untouched by this branch.

For Ben

• Flash the tip before merge if you want the post-review edge path verified on hardware; the board currently runs e3818d3.
• Merge is yours.

🤖 Generated with Claude Code

[Strycher]

Agent: EmberDelta (session f0e32f9f)