Quality pass: reuse, dead weight, and two real costs

Four review angles over the Nim port and the renderer. What got fixed:

Reuse. `summarise` and `previewLine` were the same whitespace-collapse-then-
rune-truncate function written twice, comment and all; one `summarise` lives in
`textruns` now, which is the module for string work on message text, and
`previewLine` is a call to it. `floorDiv`/`floorMod` were hand-rolled where
std/math has exactly those semantics. Every FFI entry point is resolved once
into a `final` instead of a `dlsym` plus a fresh trampoline on each call, and
the two-argument marshalling is one `_call2` beside `_call1`.

Dead weight. Nine unused imports, so the next real warning is visible. An
unused `room` parameter. `belowList` — a proc with an unused argument returning
the literal 46 — and the `reserve` prop it fed, which the renderer's own
comment says it ignores on purpose.

A line on the connect screen that had stopped being true: "TLS comes from
dart:io" is what it said, on a path where TLS is Nim's std/net. That one is
user-visible text, not a comment.

And `firstImageUrl` was ported, tested, and never called — `reducer` assigned
`imageUrl` its own default instead, so no received message ever showed the
inline preview the chat screen draws. One line.

Two costs worth the name. `currentRoom` returns a Room **by value**, and a Room
owns its whole backlog: resolving a reply called it once per replying line per
render, which in a busy room is tens of thousands of message deep-copies a
frame, ten times a second. The room is threaded down now. And `localOffsetSeconds`
was a `localtime_r` — a stat of /etc/localtime — three times per timestamp per
render; it is memoised per UTC day, which is exact rather than approximate
since the offset never moves inside a day.

The poll loop compared two trees by stringifying both, ten times a second, to
answer "did anything change". It compares the JSON Nim already produced and
decodes only on a miss.

Altitude. `hbox` + `wrap: true` + `inline: true` was three props saying
"paragraph" and a renderer branch that shared nothing with either row path and
threw away the children it had just built. It is a `paragraph` tag. And
`fillHeight`-on-a-vbox, a bare `scroll`, and a row peering at its children's
props were three spellings of "take the remaining extent" — one `expand` prop,
stated by the node that expands, handled once.

One finding I applied and reverted: flipping `hbox`'s wrap default to false,
on the grounds that the tree states `wrap` at every call site. It does not —
4 of 15 do — so the other 11 became Rows and overflowed. The layout tests
caught it, which is the second time this week they have caught something I
was about to be confident about.

Skipped, deliberately: the unused `State` fields, the `dialog` tag, the
Overview toggle and `edits.nim` are all staging for screens not yet ported —
real dead weight today, churn to delete and re-add. `frq_ui_demo` stays in the
ABI beside `frq_ui_reset`, which can wipe a session just as thoroughly. And
`clock.nim`'s Hinnant arithmetic stays: std/times could do it, but it is
tested against a fifty-thousand-day round trip and rewriting working date code
for line count is not a quality improvement.

203 Nim tests, 20 Dart, 21 layout, check-common clean, no compiler warnings,
and the window still reaches #test with 114 lines and no layout errors.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>