…per (#3239)

Completes the platform matrix promised in #3232.  Linux (evdev) landed
in #3238; this PR adds the Windows (hidapi) and macOS (IOKit) sides.
The mapper dialog drops its Linux guard and now builds everywhere.

Architecture
------------
All three backends expose the same Qt signal contract:
    tuneSteps(int)               - rotary delta
    buttonEvent(sig, action)     - canonical Linux-keycode-name strings
    connectionChanged(bool, name)

A new typedef `UlanziDialBackend` (src/core/UlanziDialBackend.h)
resolves to the right concrete class per platform.  Mapper dialog and
MainWindow dispatcher use only the alias and have no platform switches
of their own.

Windows (UlanziDialWindowsManager)
----------------------------------
- hidapi (already a project dep) enumerates HID devices, matches by
  product_string substring "Ulanzi Dial"
- Opens every matching interface (the dial enumerates as multiple
  HID interfaces on Windows: Keyboard + Consumer Control, possibly
  Mouse) and polls each via QTimer at 5ms intervals
- HID report parsing: standard 8-byte keyboard report (modifier mask
  + up-to-6-keys array, diffed across frames for transitions) and
  Consumer Control reports (2-byte usage code)
- HID usage codes mapped to canonical Linux keycode integers so the
  signature strings (`KEY_PLAYPAUSE`, `Ctrl+V`, ...) match Linux
  exactly and `kPillSpecs` is unchanged
- Chord assembly state machine identical to the Linux backend
- Hot-plug via 3s rescan timer
- LIMITATION: no exclusive-claim equivalent yet.  The dial's keystrokes
  still reach the focused window because hidapi doesn't suppress OS
  input routing.  RawInput + RIDEV_NOLEGACY is the proper fix; tracked
  for a follow-up

macOS (UlanziDialMacOSManager)
------------------------------
- IOKit HID Manager with device-matching dictionary
  ({ kIOHIDProductKey: "Ulanzi Dial" })
- `IOHIDManagerOpen` with `kIOHIDOptionsTypeSeizeDevice` -- the
  documented macOS exclusive-claim mechanism.  Equivalent to Linux
  EVIOCGRAB: when seized, the dial's events stop reaching the OS
  keyboard stack, so its media keys don't leak to the focused window
- Event-driven via `IOHIDManagerRegisterInputValueCallback`; no polling
- Same usage-page-to-Linux-keycode mapping and chord assembly as
  Windows backend
- Scheduled on `CFRunLoopGetMain()` (IOKit requires a CFRunLoop, and
  only the main thread has one integrated with Qt's event loop on
  macOS).  MainWindow conditionally skips `moveToThread` on macOS to
  match this constraint
- Hot-plug native via IOHIDManager device-matching/removal callbacks

Build wiring (CMakeLists.txt)
-----------------------------
- UlanziDialBackend.h + UlanziDialMapperDialog.{cpp,h}: unconditional
- EvdevEncoderManager.{cpp,h}: UNIX AND NOT APPLE (Linux)
- UlanziDialWindowsManager.{cpp,h}: WIN32
- UlanziDialMacOSManager.{cpp,h}: APPLE
- macOS links -framework IOKit and -framework CoreFoundation

Tested
------
- Linux: compiles + runs as before (Linux behaviour unchanged; same
  EvdevEncoderManager underneath the alias)
- Windows + macOS: compile validation gated by CI (check-windows from
  #3210 and check-macos from #3223).  Live testing on hardware happens
  in the user's environment

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

## Summary

Drop the \`check-paths\` allow-list job that gated \`check-windows\` and
\`check-macos\` based on a hand-maintained set of files. Both
cross-platform CI builds now run on every PR.

## Why

The path-filter approach has been a chronic leaker. Every time a new
file gained a platform-guarded branch (\`#ifdef Q_OS_WIN\`, \`#ifdef
Q_OS_MAC\`, \`HAVE_PIPEWIRE\`, etc.) that wasn't on the allow-list, CI
silently skipped the relevant cross-platform check and the regression
surfaced at release tag time. The follow-up was always to add another
path to the filter — and then the same class of issue would resurface
elsewhere a few PRs later.

Pattern receipts:
- #796 — original \"MSVC issue slipped past CI\" lesson
- #2633 → #2662 — macOS \`m_daxBridge\` reference inside macOS-only
guard, missed
- #2670 — Windows behavior change not exercised by CI
- #2671 — added MainWindow.cpp/.h to filter after a miss
- #2929 → #3052 — WASAPI mono-mic recovery landed unverified; added
AudioEngine.cpp/.h
- #3210 — re-added AudioEngine after a relapse
- v26.5.3 — ClientPhaseRotator.cpp \`M_PI\` MSVC break slipped through
- #3241 (yesterday) — CwSidetonePortAudioSink WASAPI fix; entire
\`#ifdef Q_OS_WIN\` block never compiled by CI because the file wasn't
on the filter

## Trade-off

Cost: ~15-20 min of GitHub Actions runner time per PR (Windows + macOS
in parallel, both cached aggressively — cold ~5 min, warm ~2-3 min).

Benefit: a whole class of release-time regression we've been paying for
repeatedly.

## Test plan

- [x] YAML valid (\`python3 -c 'import yaml; yaml.safe_load(...)'\`)
- [ ] Confirm \`check-windows\` and \`check-macos\` actually run on this
PR (no longer marked \`skipping\`)
- [ ] Both pass

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…thersdr#3210)

## Problem

`AudioEngine.cpp` and `AudioEngine.h` are as dense with `#ifdef
Q_OS_WIN` / `Q_OS_MAC` / `HAVE_PIPEWIRE` branches as `MainWindow`, but
were absent from the dorny/paths-filter `build` group that gates the
`check-windows` CI job.

A Linux-passing change inside a `#ifdef Q_OS_WIN` block can therefore
land on `main` without triggering a Windows build check. This is the
same class of issue that motivated adding `MainWindow` to the filter in
aethersdr#2671.

**Post-mortem:** PR aethersdr#2929 (WASAPI mono-mic recovery) slipped past
`check-windows` for exactly this reason — `AudioEngine.cpp` carried the
platform-guarded change but wasn't on the filter.

## Fix

Add `src/core/AudioEngine.cpp` and `src/core/AudioEngine.h` to the
`build` paths group, with a comment referencing the post-mortem. Pattern
mirrors the existing `MainWindow` entry.

Fixes aethersdr#3052

…per (aethersdr#3239)

Completes the platform matrix promised in aethersdr#3232.  Linux (evdev) landed
in aethersdr#3238; this PR adds the Windows (hidapi) and macOS (IOKit) sides.
The mapper dialog drops its Linux guard and now builds everywhere.

Architecture
------------
All three backends expose the same Qt signal contract:
    tuneSteps(int)               - rotary delta
    buttonEvent(sig, action)     - canonical Linux-keycode-name strings
    connectionChanged(bool, name)

A new typedef `UlanziDialBackend` (src/core/UlanziDialBackend.h)
resolves to the right concrete class per platform.  Mapper dialog and
MainWindow dispatcher use only the alias and have no platform switches
of their own.

Windows (UlanziDialWindowsManager)
----------------------------------
- hidapi (already a project dep) enumerates HID devices, matches by
  product_string substring "Ulanzi Dial"
- Opens every matching interface (the dial enumerates as multiple
  HID interfaces on Windows: Keyboard + Consumer Control, possibly
  Mouse) and polls each via QTimer at 5ms intervals
- HID report parsing: standard 8-byte keyboard report (modifier mask
  + up-to-6-keys array, diffed across frames for transitions) and
  Consumer Control reports (2-byte usage code)
- HID usage codes mapped to canonical Linux keycode integers so the
  signature strings (`KEY_PLAYPAUSE`, `Ctrl+V`, ...) match Linux
  exactly and `kPillSpecs` is unchanged
- Chord assembly state machine identical to the Linux backend
- Hot-plug via 3s rescan timer
- LIMITATION: no exclusive-claim equivalent yet.  The dial's keystrokes
  still reach the focused window because hidapi doesn't suppress OS
  input routing.  RawInput + RIDEV_NOLEGACY is the proper fix; tracked
  for a follow-up

macOS (UlanziDialMacOSManager)
------------------------------
- IOKit HID Manager with device-matching dictionary
  ({ kIOHIDProductKey: "Ulanzi Dial" })
- `IOHIDManagerOpen` with `kIOHIDOptionsTypeSeizeDevice` -- the
  documented macOS exclusive-claim mechanism.  Equivalent to Linux
  EVIOCGRAB: when seized, the dial's events stop reaching the OS
  keyboard stack, so its media keys don't leak to the focused window
- Event-driven via `IOHIDManagerRegisterInputValueCallback`; no polling
- Same usage-page-to-Linux-keycode mapping and chord assembly as
  Windows backend
- Scheduled on `CFRunLoopGetMain()` (IOKit requires a CFRunLoop, and
  only the main thread has one integrated with Qt's event loop on
  macOS).  MainWindow conditionally skips `moveToThread` on macOS to
  match this constraint
- Hot-plug native via IOHIDManager device-matching/removal callbacks

Build wiring (CMakeLists.txt)
-----------------------------
- UlanziDialBackend.h + UlanziDialMapperDialog.{cpp,h}: unconditional
- EvdevEncoderManager.{cpp,h}: UNIX AND NOT APPLE (Linux)
- UlanziDialWindowsManager.{cpp,h}: WIN32
- UlanziDialMacOSManager.{cpp,h}: APPLE
- macOS links -framework IOKit and -framework CoreFoundation

Tested
------
- Linux: compiles + runs as before (Linux behaviour unchanged; same
  EvdevEncoderManager underneath the alias)
- Windows + macOS: compile validation gated by CI (check-windows from
  aethersdr#3210 and check-macos from aethersdr#3223).  Live testing on hardware happens
  in the user's environment

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

## Summary

Drop the \`check-paths\` allow-list job that gated \`check-windows\` and
\`check-macos\` based on a hand-maintained set of files. Both
cross-platform CI builds now run on every PR.

## Why

The path-filter approach has been a chronic leaker. Every time a new
file gained a platform-guarded branch (\`#ifdef Q_OS_WIN\`, \`#ifdef
Q_OS_MAC\`, \`HAVE_PIPEWIRE\`, etc.) that wasn't on the allow-list, CI
silently skipped the relevant cross-platform check and the regression
surfaced at release tag time. The follow-up was always to add another
path to the filter — and then the same class of issue would resurface
elsewhere a few PRs later.

Pattern receipts:
- aethersdr#796 — original \"MSVC issue slipped past CI\" lesson
- aethersdr#2633 → aethersdr#2662 — macOS \`m_daxBridge\` reference inside macOS-only
guard, missed
- aethersdr#2670 — Windows behavior change not exercised by CI
- aethersdr#2671 — added MainWindow.cpp/.h to filter after a miss
- aethersdr#2929 → aethersdr#3052 — WASAPI mono-mic recovery landed unverified; added
AudioEngine.cpp/.h
- aethersdr#3210 — re-added AudioEngine after a relapse
- v26.5.3 — ClientPhaseRotator.cpp \`M_PI\` MSVC break slipped through
- aethersdr#3241 (yesterday) — CwSidetonePortAudioSink WASAPI fix; entire
\`#ifdef Q_OS_WIN\` block never compiled by CI because the file wasn't
on the filter

## Trade-off

Cost: ~15-20 min of GitHub Actions runner time per PR (Windows + macOS
in parallel, both cached aggressively — cold ~5 min, warm ~2-3 min).

Benefit: a whole class of release-time regression we've been paying for
repeatedly.

## Test plan

- [x] YAML valid (\`python3 -c 'import yaml; yaml.safe_load(...)'\`)
- [ ] Confirm \`check-windows\` and \`check-macos\` actually run on this
PR (no longer marked \`skipping\`)
- [ ] Both pass

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>