feat(shortcut): single/double press detection for Ctrl+G

[0xMMA]

Closes #30

Summary

Started as a 200 ms double-press detector on top of RegisterHotKey; that cannot see a second tap while Ctrl is held, so the branch replaced the input capture with a low-level keyboard hook and made shortcuts configurable.

• Windows input capture: WH_KEYBOARD_LL hook (internal/features/shortcut/service_windows.go) tracks modifier state, suppresses the matched trigger key so the foreground app never sees it, and runs the double-tap state machine on the hook thread.
• Two modes: double_tap (single tap → fix; second tap within the delay while modifiers are held → pyramidize; releasing the modifier ends the wait early) and independent (two separate bindings, each fires on keyup).
• Configurable shortcuts: settings fields shortcut_mode, shortcut_fix, shortcut_pyramidize, shortcut_double_tap_delay (100–500 ms, default 200); legacy shortcut_key is ignored (Ctrl+G was hard-wired up to v4.5.0-beta, so it never held a live value); unusable values in the file are replaced on load and refused on Save. Key-combo parser with VK mapping (keycombo.go + tests). Saved shortcuts apply immediately (settings OnSaved hook in main.go).
• Settings UI: mode toggle plus a shortcut recorder component (frontend/src/app/features/settings/shortcut-recorder/); recording pauses the hook.
• Events: shortcut:fix / shortcut:pyramidize replace shortcut:triggered; frontend exposes shortcutFix$ / shortcutPyramidize$; the silent-fix subscription moved to ShellComponent so it works on every route; Pyramidize no longer captures the clipboard on a single press.
• Clipboard: release Shift/Alt before injecting Ctrl+C/V, poll for the clipboard change instead of a fixed sleep, skip capture on the second tap.

Design: docs/superpowers/specs/2026-04-06-keyboard-hook-configurable-shortcuts-design.md (supersedes the double-press spec in the same folder).

Test plan

• Go: key-combo parser tests, go test ./internal/... (CI)
• Frontend: Vitest incl. updated subscription tests (CI)
• Windows manual, on the dev-channel build of this PR: Settings → About → Dev channel → install PR feat(shortcut): single/double press detection for Ctrl+G #31 build (prerelease v0.0.0-pr.31, always the latest CI run of this branch; Developer options are unlocked by tapping the version). Fallback: the KeyLint-windows-amd64-setup artifact of the latest CI run.
  • version string v4.4.3-beta-35-ga04faa3 — obsolete: the merge of main (2026-10-08) superseded that build; the dev-channel build carries the 2026-09-17 fixes and main's newer work.
  • single Ctrl+G in any app → silent fix, no stray "g" typed into the app, regardless of KeyLint's current route or window visibility
  • hold Ctrl, tap G twice → KeyLint comes to front on Pyramidize with the clipboard text loaded — also when KeyLint was on Fix or Settings before
  • independent mode: Ctrl+G → fix, Ctrl+Shift+G → Pyramidize
  • recorder: record a new combo, save → active immediately without restart; restart the app, combo still active; recording itself triggers no action
  • recorder: start recording, then leave Settings without pressing a key → Ctrl+G still works
  • save an invalid combo (e.g. a bare letter) → Settings shows an error, old combo keeps working
  • an old settings.json (v4.5.0-beta, only shortcut_key) loads without error and Ctrl+G works
  • 200 ms delay feels right; single tap has no perceptible lag

Update 2026-09-17

Root cause of the double-fire. SetTimer with a NULL hWnd ignores the nIDEvent argument and returns a newly generated timer ID; the code passed a constant 1 and killed the timer with that same 1, so KillTimer never matched. Because SetTimer timers are periodic, every first tap leaked a timer that kept posting WM_TIMER forever. A stray timer then fired a silent fix milliseconds into the next first tap, the second tap was treated as a fresh first tap, and the second paste landed after the first paste's collapsed caret — so hold-Ctrl double-tap ran fix twice and Pyramidize was unreachable.

What changed.

• fix(shortcut): kill the double-tap timer by its real Win32 ID (#30) — the service now stores the ID SetTimer returns, routes every kill site through a stopTimer() helper (keyup, modifier release, second tap, UpdateConfig, SetPaused, Unregister), ignores WM_TIMER whose wParam is not the current ID, and kills the timer when it expires. If SetTimer ever fails it logs a warning and degrades to a plain single-tap fix instead of swallowing the key. The bogus doubleTapTimerID constant is gone.
• fix(shell): ignore shortcut while a silent fix is in flight (#30) — ShellComponent.silentFix() now guards against re-entry, so two fix events can no longer run two concurrent read → enhance → write → paste cycles and paste on top of each other. Covered by a new Vitest case (138 tests green).
• fix(shortcut): reset double-tap state on the pump thread (#30) — a SetTimer(NULL, …) timer belongs to the thread that created it, so the KillTimer calls made from Unregister, UpdateConfig and SetPaused on their own goroutines could never kill it; those callers now post a wmReset message and the message loop does the reset on the pump thread, which also removes the cross-goroutine writes to tapState/indepPending.
• fix(shell): release silent-fix guard after a safety timeout (#30) — a 120 s safety timer releases silentFixInFlight if the enhance call never settles (the backend HTTP timeout lands in feat(provider): Claude Code CLI as first-class provider (claude -p) #32), plus reject-path and timeout-path tests (140 green); on the Windows side a failed SetTimer now stays in tapWaitRelease so auto-repeat cannot fire one fix per repeat.
• fix(shortcut): absorb held trigger key after the double-tap window expires (#30) — holding the shortcut down used to fire one fix per delay interval, because the expiring timer dropped the state machine to tapIdle and every auto-repeat keydown read as a fresh first tap; a new tapFiredWaitRelease phase fires the fix once and swallows the trigger key until its keyup.

Verification. GOOS=windows go vet ./... clean apart from three pre-existing possible misuse of unsafe.Pointer warnings, Windows cross-compile with mingw succeeds, go test ./internal/... and cd frontend && npm test green on Linux.

Manual checklist reset. All Windows manual items above are unchecked again and must be re-run against the artifact from the CI run for a04faa3 — no earlier artifact carries this fix. (2026-10-08: use the dev-channel build instead, see the test plan.)

Update 2026-10-08

Merged origin/main (settings atomic writes + settings.Update, provider switching #104, dev channel #107, shell footer #106, Go 1.27, CI apt fixes). Conflicts in shell, wails.service, settings imports and bindings; both sides kept. Then fixed what the merge review found:

• fix(shortcut): apply saved shortcuts at once, refuse unusable ones — nothing ever emitted settings:changed, so the hot-reload never ran. Settings now calls an OnSaved hook after every write; main.go re-applies shortcuts when they changed (and installs the hook if startup registration failed). Save refuses an unusable shortcut change; load() repairs unusable values; the shortcut_key migration is gone (it never fired, and firing it would have applied unvalidated free text).
• fix(enhance): load clipboard for a shortcut sent from another page — Pyramidize by shortcut from any page but /enhance opened empty: the event fired before the page subscribed. WailsService now keeps it pending until a page takes it.
• fix(settings): resume shortcuts when the recorder goes away — leaving Settings mid-recording left the hook paused until restart; Shift+digit records the digit.

🤖 Generated with Claude Code