Unlock latency

Firmware and helper 0.1.34-dev.1 remove several waits between touching the sensor and typing a password. Install both parts to get the full improvement. No fingerprint thresholds, authentication checks, replay protections, or lift requirements change.

Changes

Stage	Previous behavior	Current behavior
Detect touch	Check the GPIO every 10 ms	A GPIO edge wakes the task; the level and lift state still gate capture
Read sensor reply	Wait to fill a 96-byte buffer, even for a short ACK	Wait for one byte, then drain available data; deliver short packets after two idle UART symbols
Show match result	Finish the LED command and wait 350 ms before proceeding	Show and clear the result in a separate task while authentication continues
Read helper request	read(256) can delay a complete short frame until the 200 ms timeout	Read available bytes immediately; block for one byte when idle
Process larger requests	Sleep 10 ms between received batches	Process buffered batches consecutively
Translate keyboard layout	Rebuild the map on every touch	Cache translations by layout contents, checking the current layout every time
Receive helper reply	Poll CDC input every 10 ms	Wake the console when CDC data arrives
Type password	USB poll interval and readiness polling both 10 ms	Request 1 ms USB polling and advance on completed reports

The CDC transmit buffer also holds a complete multi-host request, reducing waits while it is queued. LED feedback still returns to the configured idle state after the configured feedback duration (350 ms by default) even when the helper is absent. Foreground enrollment cancels older feedback.

Typing speed

New configurations default to a 1 ms minimum after each key press and release. Existing configurations keep their stored value; select the faster setting with:

tinytouch config typing_delay_ms 1

The old 100 Hz scheduler rounded settings below 10 ms down to zero. The current implementation uses a timer to honor millisecond settings and waits for actual USB completion, including the final release. Repeated characters retain separate press/release reports. A transfer failure stops typing and prevents a later Enter; if the transport is still available, it attempts to release any pressed key.

A 20-character password plus Enter needs 42 reports. With a simulated host completing each report in 1 ms, the transport takes 42 ms at a 1 ms setting, or 294 ms at 7 ms. These are model timings, not measured Mac login times. Host scheduling, application behavior, fingerprint matching and macOS login processing add their own latency. Increase the delay if a field drops characters.

Measure on a device

After an ordinary unlock, use tinytouch logs. Each entry has a monotonic millisecond timestamp:

Difference	What it measures
touch_detected → finger_matched	Sensor capture and matching
finger_matched → hid_requested	Preparing the authenticated request
hid_requested → hid_response	CDC transmission and helper response
hid_response → hid_typing	Queue delivery and response verification
hid_typing → hid_typed	HID delivery and configured pacing, through final release

hid_response means a response arrived; hid_typing follows successful verification. Failures end in hid_failed. A legacy multi-host fallback is marked hid_legacy_retry. Logs contain no passwords or key material. The 32-entry ring holds only recent activity, so retrieve it shortly after a test.

Compare several first unlocks after connection, steady-state unlocks, and wake from sleep. Check the actual login field as well as logs: completed USB transfers do not prove the application accepted the text. Use a disposable test password with repeated letters and shifted characters, test with Enter disabled in a visible field first, and test both your normal layout and any layout you switch to. Check that a held finger types only once and a missing helper does not leave the LED green.

Reproduce the host framing benchmark

From a checkout with pyserial installed:

python3 scripts/benchmark-serial.py --trials 20

This uses local pseudo terminals and synthetic EV/EV2 frames; it opens no physical device and sends no password. It compares the previous read/sleep loop with the current reader and reports median and 95th-percentile framing times. The request is queued immediately before reading, which exposes nearly the full old timeout. Actual touch timing can land anywhere within the old 200 ms read window.

Physical ESP32-S3/sensor and macOS acceptance testing is still required. The host tests cover framing, authenticated replies, layout cache invalidation, sensor packet validation, LED expiry, report pacing, repeated keys, disconnects and late USB completions. They cannot measure the sensor's recognition time or macOS's unlock time.