* feat(render): capture terminal width at the wire boundary (kz8.4)

Eliminates the per-render `ps` → `stty` → `tput` cascade in
src/utils/terminal-width.ts (up to 3 spawns per render inside the
daemon, every render).

The shape: terminal width is data flowing in from the live client's
shell context — captured once by the Rust client (env $COLUMNS, then
TIOCGWINSZ ioctl on stderr) or the Node fallback (env, then
process.stdout.columns), threaded through the wire protocol as
RenderRequest.termCols, and consumed unchanged by the renderer. The
daemon's own env is meaningless here (it reflects whichever shell
launched the daemon hours ago), so only the client can answer.

`src/utils/terminal-width.ts` collapses from 117 lines (4 spawn sites)
to 33 lines of pure resolution: hint → env → stdout → null. Dead code
(getRawTerminalWidth, Windows mode con, parent-tty walking) is gone.
"terminal-width" is removed from LAUNCH_CATEGORIES so the metering
surface shrinks with the spawn surface.

Wire protocol bumps to v3 (TS + Rust in lockstep). Existing daemons
hit VERSION_MISMATCH and respawn from the new binary — the documented
restart path.

[LAW:single-enforcer] One source for terminal width: the client.
[LAW:dataflow-not-control-flow] Renderer runs the same code path
every time; the data (termCols set or not) decides what happens.

Tests: +7 terminal-width unit tests (purity, hint precedence, env
fallback, malformed input), +1 render-request integration test
proving termCols actually reaches the autoWrap layer. 704 tests pass.

* gitignore

* docs(terminal-width): describe what the module IS, not what it once WAS

"This module no longer spawns subprocesses" reads against a past state
that the future reader doesn't have context for. Replace with the
timeless shape ("is a pure resolver", "subprocess fallbacks belong at
the wire boundary") so the comment stays true regardless of when it's
read.

…ep (1qn) (#102)

Sheriff finding #9 [LAW:no-ambient-temporal-coupling]: handleRequest
scheduled setTimeout(() => shutdown(0), 50) after a newer-client version
mismatch and after the shutdown verb, encoding "the response has flushed
by then" as a magic delay nobody owned. A slow flush meant the client saw
a dead socket (transient/io_error) instead of the VERSION_MISMATCH
diagnostic (permanent/version_mismatch) — exactly during upgrades.

The flush completion now owns the ordering: handleRequest returns
{ resp, exitAfterFlush: number | null } — the exit wish as data
[LAW:effects-at-boundaries] — and respond() at the socket boundary
performs it via sock.end(frame, cb), whose callback fires on finish OR
error (total signal; a vanished peer cannot strand the exit). The
version-mismatch asymmetry is now literally the value
`req.v > PROTOCOL_VERSION ? 0 : null` [LAW:dataflow-not-control-flow].
The SIGKILL backstop inside shutdown() is untouched.

New integration test pins the contract: a v+1 client receives the flushed
VERSION_MISMATCH frame, then the daemon exits within budget.

Co-authored-by: elton-prawn <264370887+elton-prawn@users.noreply.github.com>