Rebuild the studio's fixed chrome so the mixer gets the room

[thcp]

Four problems in the studio's fixed chrome, all found while measuring where the
mixer's vertical budget goes.

Closes #487
Closes #488
Closes #489
Closes #490

Measured, at 1366x768

The resolution #480 was reported at, library already collapsed.

	Before	After
Analysis header	147 px	73 px
Footer	212 px	170 px
Mixer	~200 px	376 px
Stem lanes fully visible	2	5 of 6

#480 stays open. The mixer still scrolls at six stems, and that issue asks
for a way to reduce these panels rather than for them to be smaller once. This
is the cheapest part of its budget won back, not the answer to it.

The analysis header (#487)

Each metadata card was three stacked text lines. The qualifier now sits on the
value's own line, E min Natural Minor, and the presence cards read
VOCAL PRESENCE 66% on one line rather than stacked. Type came down with it,
22 px to 17 px and 28 px to 18 px.

Halving it was a question of how many lines a card is, not how big its text is.
Three stacked lines have a floor: 10 px labels are already at the bottom of what
is readable.

The header's two rows (#488)

Seven metadata cards sat above six presence cards, each unaware of the other, so
their dividers landed at different x all the way across. Both rows are now grids
of the same seven columns. Six presence stems fill seven because Vocals takes
two; once the lead/backing split has run it is replaced by two cards and the row
becomes a plain 7x1. Verified: every bottom edge lands on a top edge, in both
states.

The top row had to become a grid too. Flex and grid round fractional track widths
differently, and as a flex row it was a pixel out of step at most widths.

Those two new cards had nothing to show. split_vocals never touched
stem_presence, so after a split neither half had a figure anywhere. Presence is
a stem's RMS as a percentage of the loudest stem in the job, and that reference
is not stored: only the percentages are. The obvious fix decodes all eight stems
again to place two of them on a scale the other six already sit on.
presence_for_split recovers the reference arithmetically instead:

loudest = rms[vocals] / (presence[vocals] / 100)

merge_stem_peaks already measured RMS for the new stems in its existing pass
and discarded it. It returns it now, with vocals riding along in the same pass:
one extra file scanned. With no reference recorded it returns nothing and the
cards read --, which is honest. Three tests cover it, including the empty case.

The footer's controls (#489)

The clusters sat in a wrapping flex row, so the click track either shared the
line, took its own, or stayed and split its own options across two rows. The
footer was never the same shape twice.

Nothing wraps and nothing shrinks now. Verified at 1920, 1600, 1280 and 1024:
one row, order fixed at transport, position, speed, key, click, every cluster
34 px tall. A window too narrow to hold the row scrolls it sideways, on the same
thin auto-hide scrollbar as the library list and mixer.

The row is top-aligned, so every label sits on one line no matter how tall the
controls under it are. Bottom-aligned, the click track threw its own label a row
above the other four.

That also settled the dividers. They were hidden when stranded at the end of a
wrapped line, but bottom-alignment meant the check never matched and none were
ever drawn. Nothing can be stranded now, so syncFooterDividers and the
is-orphan rule are deleted rather than left as logic that cannot fire.

min-height: 200px dated from when the click cluster could wrap. Its content
needs 170, so the floor only padded the footer past itself and the surplus
pooled as a gap under whichever tier came last.

The track identity (#490)

It was holding a 300 px column of the footer to answer a different question from
everything beside it: the footer's left column said what is loaded, the rest said
how it plays.

It went to the library sidebar first, and came back. The gutter beside the
timeline has to exist regardless, because the timeline's left edge lands on the
lane waveforms' left edge so a position reads at the same x in both. Leaving the
identity in the sidebar meant a panel whose height had to be measured and
mirrored against the footer's, and it put the loaded track's name in the column
you read to choose a different one.

So it is back in the gutter, but in the tier that gutter belongs to rather than
as a full-height column beside the controls. The controls keep the left edge they
gained, and the block is one tier tall instead of two: cover and title on a line
with the favourite, the meta line under them. 79 px, against 141 in the sidebar.

Export Mix moved to the right of the controls row. It acts on the track rather
than on playback, so it sits opposite the clusters instead of filed among them,
and its menu opens toward the middle of the window: anchored left, a 260 px menu
on the footer's right edge would hang off it.

That placement needed the row split. The clusters scroll sideways when the window
cannot hold them, and anything after them would scroll off with them, so
.footer-clusters takes the overflow and the actions sit outside it. Export Mix
is reachable at any width.

Its footer-* classes travelled as track-panel-* and track-chip-* while it
was in the sidebar, and kept those names: they describe the block, not where it
sits. Every id is unchanged, so no JS or test needed touching beyond one selector
in main.js.

Measured: lane panel and footer timeline both at x=690; identity slot 390 to 690;
Export Mix right edge 1582, flush with the timeline's.

The rail reads Favorites, Library, Queue, Trash. Queue was below Trash, an
active state filed under an archive.

Verification

66 browser tests pass, including the whole export-menu suite, which exercises the
export button in its new home. 845 backend tests pass; the 14 failures are
pre-existing and reproduce identically on main, nine of them because this
machine has no ffmpeg. ruff check and ruff format --check clean, node --check clean on every file. i18n complete at 468 keys across all nine tables:
presence.vocalGlobal added, presence.vocal removed.

Before merging

• uv.lock is untouched, so the desktop in-app updater is unaffected.
• The Unraid template still pins 0.15.2.