feat: show the input name and downtime on the split-flap board (#154)

[BernardJen]

Closes #154.

The board cycled five hardcoded word triples. It now opens on what the operator actually
wants to know:

HDMI 01
NO SIGNAL
DOWN 04:31

The contract decision, which #154 said was the real work

Option 3 of the three it weighed: the renderer pushes rows in via a new setMessages(),
and the board stays a dumb renderer of whatever it is handed. All knowledge of inputs, sides
and downtime stays in renderer.js.

Chosen over passing app state into create() because this board is not a registry entry —
renderer.js imports it directly — so nothing had to change about the saver contract to do it.
The weather saver (#101) answered the same question the other way for a different reason: it
needs state the renderer does not own, so it reads a cached value from its own module. Here the
renderer owns everything, so pushing is strictly less coupling.

Rows apply on the board's next message change, not immediately, so a push never interrupts
a flap in progress. With a single-entry list the board re-flaps the same rows each hold period —
which is what makes a ticking clock read as a mechanism rather than a redraw.

Truncation, not a wider grid

MIN_COLS stays measured over the built-in messages only. Growing cols at runtime would mean
reallocating the tile arrays and the state texture mid-flight; #154 explicitly allows
truncation, so the cheap option is also the sanctioned one.

Computed from the board's own formula:

View	Per-board	Columns
Dual (one side)	3000×1200	16
Single (full wall)	6000×1200	22

An input named longer than that truncates — BLACKMAGIC ULTR. — and renaming it in settings is
the fix. The cut ends in a period rather than clipping silently, because silent clipping is the
exact failure #92 fixed: NO SIGNAL rendered as NO SIGNA reads as a broken renderer, not as
an abbreviation.

The clock has three forms so it reads correctly at both ends — DOWN 04:31, DOWN 2H 14M,
DOWN 3D 04H. The colon form only ever means minutes and seconds, so it is never ambiguous
against the hour form. A test asserts every character it can emit is in FLAP_CHARS.

Two things that would have failed quietly

A repeated showNoSignal must not restart the clock. It is idempotent and gets called again
for a synced same-device pair and on every re-entry, so restarting would peg the downtime line
near zero forever — looking like it works while never advancing.

A malformed push must not blank the board. setMessages validates and falls back to the
built-in list, so a renderer with nothing to say leaves the static messages rather than a grid
of spaces.

Tests

23 new, 595 → 618. Each invariant confirmed by breaking it:

Mutation	Caught
restart the clock on every showNoSignal	✅
clip without a marker	✅
drop the push validation	✅

Verified, and not

Driven in a browser with both sides down, one backdated four minutes and the other two hours.
Both boards started, both took live rows, and the clock advanced across the run. Fitted output
at the real column count: HDMI 01 / NO SIGNAL / DOWN 06:05 and
BLACKMAGIC ULTR. / NO SIGNAL / DOWN 2H 15M.

Not verified: the flaps settling on that text. The Browser pane throttles
requestAnimationFrame, so the board crawls mid-flip and never lands — the screenshots show it
animating toward the new rows rather than displaying them. Rendering arbitrary strings through
this path has been shipping since #147, and only the strings changed here.

One correction worth recording: my first in-page column derivation used TILE_RATIO = 0.72
dividing, when it is 1.45 multiplying. Both produced 16 for that canvas because MIN_COLS
bound either way, so the fitted strings above are right — but the table above comes from the
real constant, not that one.