shurectl QA pass

shurectl --qa runs the real app (handle_key(), mouse::handle_mouse(), apply_action(), ShureDevice) against a fake device and renders into a buffer instead of a terminal. Each line you send is answered with one JSON line. Nothing touches real hardware or ~/.config/shurectl.

cargo build --features qa
./target/debug/shurectl --qa --fake-device mv7plus   # packets logged by name
./target/debug/shurectl --qa --demo mv6              # no device: navigation only
printf '{"key":"Tab"}\n{"query":"screen"}\n' | ./target/debug/shurectl --qa --demo mv6

Models: mvx2u, mvx2u-gen2, mv6, mv6-gen2, mv7, mv7plus. The fake device holds Manual mode at 24 dB, read back at startup as from a real device. Demo mode starts where shurectl --demo does: DeviceState::default(), which is Auto Level at 36 dB. Run a pass once per model; a bug on one model is usually a divergence from the others.

Lines you send

Inputs:

Line	Does
{"key": "Tab"}, {"key": "BackTab", "mods": "shift"}, {"key": "c", "mods": "ctrl"}	One key press, named as crossterm names it (Enter, Esc, Up, Left, Backspace, Space, F1, or one character). "repeat": n presses it n times
{"type": "My Preset"}	One key press per character
{"click": [col, row]}, {"wheel": "up", "at": [col, row]}	Mouse at a cell
{"click_on": "Mute:"}	Click the first cell of the first row containing that text
{"resize": [80, 24]}	Redraw at a new size (max 300x120)
{"tick": 2100}	Advance the fake clock (ms, max 60000); runs the presence poll (every 2 s) and the mute poll (every 1 s, a mute GET or micMute on the MV7) when they fall due
{"device": "unplug"}, {"device": "replug"}	Fake device only
{"device": {"drop_acks": 1, "extra_acks": 1, "refuse": 1, "fail_io": true}}	Inject faults: fewer or more acks after the next CONFIRM, answer the next command with an unknown address ([Failed] on the MV7), fail all I/O. A tick that runs the mute poll can use up refuse, and its drain discards queued acks, so put a one-shot fault right before the command it is for
{"device": {"muted": true}}	Mute (or with false unmute) the fake on the device itself, as its mute button does; the next mute poll shows it
{"meter": -12.5}	Push a peak (dBFS) into the meter at the fake clock's time; tick ages it (bar 0.3 s, peak hold 3 s). Demo mode has no meter
{"preset_dir": "fixtures/presets-v2.9"}	Copy preset files into this session's temp preset dir
{"reset": true}	Start over on the same model

Queries (answer in result): {"query": "screen"}, "screen_styled" (runs with palette names like C_ERROR), "state" (every DeviceState field), "hitmap" (every click region: use it to find targets on a squashed terminal), "packets" (the whole log), "focus_ring" (what Down visits from here), "models", "features" (names the fake decodes).

Every reply has ok, app (tab, focus, status, status_is_error, help_visible, editing, connected, quit, fits, …), screen, changed (rows that differ from the previous reply), packets since the previous line, failures and errors. A packet the fake refuses (malformed, not something protocol.rs builds) is an error: report it.

Expectations

{"expect": {...}} checks every key and reports the actual value on failure; exploring continues.

Key	Checks
screen_contains, screen_not_contains (string or list), screen_row: [n, "text"]	Screen text. Prefer screen_contains over row numbers
screen_hash	The renders_match_golden hash, for a pinned screen
style: {"text": "Muted", "fg": "C_ERROR"}	Color of the text's first cell, by palette name
tab, focus, status (substring), status_is_error, help_visible, editing, quit, confirming_factory_reset	App fields
fits	ui::fits() for the last frame: false below 80x24 (MIN_SIZE), where only the size notice draws and keys only quit
connected	App::device_connected, and the header agrees
state: {"muted": true}	DeviceState fields (see the state query for names)
set: {"feature": "GAIN", "value": 250} or set: "GAIN"	A SET since the previous expect line. "applied": false means the device ignored it
sets: [...]	SETs in this order (others may come between)
get: "SERIAL", text_sent: "micMute on" (MV7), no_packets: true, packets_raw: ["01 12 11 22"]	Other packets since the previous expect
click_targets: true	Clicks every unfocused control region: each must focus that control and send nothing

In demo mode the packet keys are skipped, not failed.

A QA pass

Read Cross-Device UI Consistency in AGENTS.md first. Then, for every model:

1. Walk every tab with Tab and BackTab, and every focus with Up/Down (focus_ring).
2. On each focus, press Left, Right and Enter, and check the status message, the screen and the packets. Every visible, focusable control must do something.
3. Repeat the Main tab at heights 24, 30, 40 and 50 with click_targets. Below 80x24 only a notice draws (terminal-too-small).
4. Compare models: the same control must have the same label, units, formatting, key, status wording and position relative to its neighbours. For each shared control check its label and hint ([Enter] toggle), value format (24 dB, +2.0 dB, 75 Hz), status after the same action (Mute → ON), the keys that act on it and that no other key does, its place in the focus ring and in the drawn block order, its colour per state (screen_styled), its Summary row, and the wording for shared conditions (disconnected, reconnected, refreshed, empty slot, refused reply). Compare the pairs first (MVX2U Gen 1 and Gen 2, MV6 and MV6 Gen 2, MV7 and MV7+), then everything against the MV7+, which has the most controls. The intended differences listed in AGENTS.md are not findings.

Report each divergence as a recorded scenario whose expect fails today. Classes worth looking for: a focusable control that does nothing on one model, a label or status spelled differently on two models, a click region on the wrong control at one height, a key only one model answers, a packet sent in demo mode.

Saving a scenario

{"record": "start"}
{"key": "Down"}
{"key": "Enter"}
{"expect": {"set": {"feature": "MUTE", "value": true}}}
{"record": "save", "as": "tests/scenarios/mute-on-mv6.jsonl", "models": ["mv6"]}

Paths are relative to where --qa runs (tests/ under the end-to-end runner). Trim the file to the lines that matter. cargo test --features qa --test e2e runs every file in tests/scenarios/ per model in its header and in both fake and demo mode (add "modes": ["fake"] when it needs the device or the fake's Manual, 24 dB start). The runner refuses a file with no expect line.

A bug you can't fix in the same change can still go in as a regression test: write the expect lines for the intended behaviour and add "known_bug": "what is wrong" to the header. The runner then requires an expectation to fail, and fails once the scenario passes, so the fix removes the marker.

A recording asserts what the app did, which may be the bug. In the PR body, say for each expect line whether it states intended behaviour or only observed behaviour.

Limits

The fake is built from protocol.rs, so it agrees with the app on every address and encoding. It catches sequencing, routing, UI and state bugs, never a protocol misreading; hardware truth is usbmon captures. It does not emulate the MV7's help listing (that needs a capture from a real MV7) or the side effects of an MV7 mode change on EQ and compressor.