Transpose the whole track, or one lane at a time, on one clock

[thcp]

Transposition, per-lane and global, plus two things found on the way.

Closes #245
Closes #476
Closes #477
Closes #478

Transpose (#245, #476)

A semitone control in the transport bar moves the whole track. One on
every mixer lane moves that instrument alone, so a harmony line can be
practised against the original melody rather than instead of it.

Drums are never transposed, and that is enforced in three places
because there are three ways in: the UI never offers it, the engine
ignores it if a caller asks anyway, and the worklet's unpitched input is
where drums are wired regardless. It also covers drums you cannot see.
A stem deselected at import is folded into original.wav, so
playbackStems.js rebuilds that lane from the retained raw stems and
routes its drums out. A job too old to rebuild keeps the whole lane
unpitched rather than risk resampling a kit mixed into it.

One clock, not one shifter per lane

The obvious build does not work. WSOLA latency moves with the shift, so
independently clocked chains put lanes up to 144 ms apart:

pitch  0, tempo 1       0 ms   (bypass)
pitch +2, tempo 1     115 ms
pitch -5, tempo 1     144 ms
pitch  0, tempo 0.75  437 ms

So the worklet takes one input per semitone, each a pitch chain, summed
into a single tempo stage. A lane's key is which input it is connected
to, not a parameter. Measured with five chains live at once:

Worst pitch error	0.35 cents
Worst skew between buses	12 ms
Drift over 60 s	0.0 ms
Semitone 0	bit-identical bypass

Changing a lane mid-playback neither restarts nor flushes. Every bus is
primed with the same silence, so the old chain plays out what it
buffered while the new one takes over as it runs dry. A short duck, not
a gap.

Absolute lane keys

A lane shows the key it is actually in, and the global control applies
its change rather than its value. So a lane deliberately put a third
above the rest stays a third above the rest when the track moves, and
there is never a second number to add in your head. Reset returns the
global key and every lane at once, which stepping the global back to
zero cannot do once lanes have moved individually.

Three things the measurements changed

• An earlier revision split attacks out of the music bus and passed them
  at unity while the rest went through the pitch chain. A delayed copy
  plus an undelayed copy cannot sum back to the input: it put a 0.64
  full-scale step into the output at every onset, against a source whose
  own largest step was 0.10. Heard as static. Removed.
• The chain produced exactly one quantum per quantum with no slack, so a
  quantum occasionally came up short and was emitted zero-padded. A hard
  edge to zero mid-block is a click. Fixed with an output reserve and an
  input cushion banked during priming.
• The old alignment test reported 2 ms, but only because the attack
  splitter was passing its probe through the bypass. Real alignment was
  18 ms and one-sided, being the WSOLA sequence sawtooth. Predictable,
  so now cancelled rather than tolerated.

Lane meters (#477)

They were linear in amplitude. Loudness is logarithmic and a separated
stem sits around -20 to -30 dBFS, so normal playback lived in the bottom
tenth of the bar. Measured median across a real library: 10.4%, with
drums and "other" under 2%.

Now dB-scaled with a -60 dBFS floor:

stem (median)	before	after
drums, quiet mix	1.7%	27.8%
other, quiet mix	1.6%	26.9%
vocals	5.0%	43.4%
bass	34.1%	71.1%

This forced the meters off getByteTimeDomainData: its quantisation
step, one part in 128, is -42 dBFS, which on a dB scale would light
every meter to a third of full over silence.

Windows test harness (#478)

tests/e2e/serve.sh is stored LF with no .gitattributes, so a
Windows checkout wrote it CRLF and bash could not run a single line of
it. The browser suite was unrunnable on the primary desktop development
platform, and CI never saw it because CI is Linux.

Testing

suite
Browser (Playwright)	66 pass
DSP	21 pass
Engine routing	43 pass
VU scale	14 pass
Other JS suites	pass
i18n coverage and duplicates	clean, 9 tables plus the ptPT variant

Two of these deserve a note. tests/js/pitch-shift.test.mjs includes a
control that rebuilds the worklet with the input cushion removed and
confirms the click gate fails, so it can actually catch the regression
it exists for. tests/js/vu-scale.test.mjs asserts what fraction of the
bar real measured stem levels light and keeps the old linear formula as
a control, so the comparison is a measurement rather than a claim.

Not covered here

tests/e2e/serve.sh runs uv run, which resynced a project environment
on a GPU box, uninstalled seven packages and tried to replace a
hand-installed torch 2.6.0+cu124 with the plain wheel. It failed
partway on a locked DLL and rolled back, so nothing was lost. Noted in
#478 as worth its own issue; not fixed here.

No Python dependencies changed, so the desktop in-app updater is
unaffected.

Before merging: the Unraid template pin in templates/stemdeck.xml is
still 0.15.2 and is a separate decision.