feat: snapshot thumbnails on the dropdown input rows (#242)

[BernardJen]

Closes #242.

Each input in the dropdown now shows a still of what it is sending, taken when the dropdown
opens.

Snapshots, not live previews — which is what made this small

The app holds exactly two streams. Four live previews would have meant rebuilding that; for
a picker open a few seconds it buys nothing. Iterating means at most one extra stream open
at any moment, so this needed no change to how feeds are held — and #261, the N-feed
refactor, closed unbuilt.

Where each still comes from is the point:

Input	Source
already on screen	straight off the live <video> — no stream, no negotiation
not on screen	a temporary getUserMedia, released the moment a frame is drawn

A cold capture device takes a visible fraction of a second to negotiate, so the sweep is
sequential and rows fill in as stills land. The tile renders before any still exists and is
its own placeholder, so a row never changes height. A device that never presents a frame
times out and stays a placeholder — which is also what an exclusive-access device does, and
what an input with nothing plugged in looks like. A failed capture is a normal outcome
here, not an error.

The first layout was wrong, and measuring caught it

Full-row-width tiles (width: 100% + aspect-ratio) measured at a 1280×800 window:

List	Tile	Row	Panel content vs cap
single	292×164	218px	1166px vs 640px

A four-input picker that scrolls is absurd for something you hover over to choose an input.

A fixed 128×72 is enough to recognise an input, and it changes what fits: the name sits
beside the tile in the single list (~322px rows) and below it in a dual column (~149px).
Two rules, each matching its width.

After, at the wall's 1200px height, neither mode scrolls — 881px and 678px against a 960px
cap.

Verified in both modes this time. Checking only Single is exactly what I got wrong on
#258 and what produced the bug you reported against it.

Other things worth knowing

Mock inputs need their own branch. getUserMedia rejects a mock-input-N id, so
openTemporaryStream calls createMockStream instead. Without it the feature would be
untestable in exactly the mode built for testing it (#248).

Rows are found by [data-device-id], not held references.
renderDropdownInputLists() rebuilds the row DOM on any device or selection change, which
can happen mid-sweep. Stills cache as data URLs (~4–6KB each), so re-opening shows the
previous still immediately while a fresh sweep replaces it, rather than flashing back to
placeholders.

Both open paths trigger it — mouseenter on the trigger (a listener that already
existed for the cursor) and toggleDropdown for touch. The panel's visibility is pure CSS,
so that listener is the only JS signal a hover-open happened.

Tests

23 new, 559 → 582. Verified by breaking each invariant:

Mutation	Tests failed
parallelise the sweep	3
leak the stream (no stop)	4
drop the re-entry guard	1
stretch instead of crop	1
tile back to full row width	1

One test bug fixed on the way. Two tests replaced the shared canvas stub and never
restored it, so every later test in the file ran against a context that threw. The cache
test was failing for that reason, not for anything in the code — worth saying because a
suite that poisons itself is worse than one test short.

Not verified

Real capture hardware. Every still here came from a mock input or a live <video>; how long
an actual capture card takes to negotiate, and whether any are exclusive-access on the
wall's host, is what FRAME_TIMEOUT_MS (2500ms) is guessing at. That number is the one to
adjust if tiles come up empty on real inputs.