…3644)

## Summary
Fixes #3623 — the CWX local sidetone drops/clips individual CW elements
under panadapter/VITA-49 load on Windows.

## Root cause
`CwxLocalKeyer` scheduled every element edge with a `QTimer` parented to
`MainWindow`, so the timer ran on the GUI event loop and
`keyStateChanged` was delivered through a GUI-thread slot. Under
panadapter paint + VITA-49 burst handling the loop coalesces both, the
key gate lands a block or two late, and an element gets gapped/clipped.
#3202's drift-correction fixes *cumulative* slip (the message stays in
sync) but not this *per-edge* jitter, and reasons specifically about
macOS — this is the residual Windows case.

## Fix
Re-architect `CwxLocalKeyer` to mirror **`IambicKeyer`**, which already
abandoned `QTimer` for exactly this reason (*"QTimer's jitter is too
high for CW"*):

- Run the element schedule on a dedicated worker thread.
- Time each edge to an absolute `std::chrono::steady_clock` deadline
waited on an interruptible `condition_variable` — no QTimer, no event
loop.
- Drive the sidetone gate via a lock-free `onKeyDownChange` callback;
`CwSidetoneGenerator::setKeyDown` is `std::atomic`, the same gate path
the iambic keyer uses.

`encode()` and the drift-correction logic are unchanged, so message
timing and the live-mode CharGap behaviour (#2473/#2754) are preserved.
`MainWindow` now owns the keyer as a `unique_ptr` and stops it (joining
the worker) before `AudioEngine` is torn down, since the gate callback
touches `m_audio->cwSidetone()`.

This is the approach recommended in the issue triage.

## Testing
- **`cwx_local_keyer_drift_test` (#3271)** retargeted to the new clock
seam (worker-disabled synchronous subclass + callback) — **passes**; all
drift-correction assertions (absolute edge targets, slip-shortened
waits, sustained-overload clamping, queued-macro epoch continuity,
stop/drain resets) still hold, so the scheduling math is unchanged.
- Clean build (Windows MSVC, Qt 6.10.3).
- Ran the app against a synthetic FlexRadio source under a heavy
spectrum load and sent long CWX macros: **CW sends cleanly at both 20
and 40 wpm — the element drops/stutter are gone**, including at 40 wpm
where the tighter element timing is the more demanding case. Clean
startup / connect / shutdown (worker joins without hanging).

## Known residual
Verified clean at **both 20 and 40 wpm** under heavy synthetic load. A
single faint, low-level click heard at 20 wpm did not recur and isn't
speed-correlated (40 wpm — tighter timing — is completely clean),
consistent with an incidental audio under-run under the deliberately
brutal synthetic load rather than the keyer. `CwSidetoneGenerator` is
unchanged by this PR and its raised-cosine envelope ramps cleanly across
block-quantized edges. The "fullest fix" floated in the issue (advancing
the gate by sample count inside `CwSidetoneGenerator::process()`)
remains available as a follow-up if sub-block edge precision is ever
wanted.

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

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>