First-run onboarding keeps asking after /quit, and never says the answers were dropped

[wine-fall]

What it is

Quitting mid-onboarding is specified to discard everything (spec 06 §2.1
"Interruption"), but the way it plays out reads as a lost answer rather than a
deliberate discard, and the loop does not actually stop when the listener
leaves.

From .dev/dev.log, a run against a fresh MURMUR_HOME (2026-08-26):

13:04:52 host: This is murmur's first run ... Three short questions ...
13:04:52 host: Who's listening? ...
13:04:58 tui:  command received: /quit
13:04:58 host: What do you want on the air? ...
13:04:58 host: How do you like to be talked to? ...
13:04:58 host: stopped before the broadcast.

Two things, one seam:

1. The questions keep coming after the listener has left. runFirstRun's
   loop (src/first-run.ts:110-113) calls ask() for all three questions
   unconditionally; only the read short-circuits, because the quit latch
   resolves every later read to ''. So Q2 and Q3 are put to someone who has
   already quit — both landed in the same second as the /quit above. The
   quit check is after the loop (src/first-run.ts:118), not inside it.

2. Nothing says the answer was dropped. The listener typed a real answer to
   Q1, then quit, and the only closing word is "stopped before the broadcast."
   The next boot starts again from Q1 — correct per spec, but indistinguishable
   from murmur having failed to remember. This is what the report was: "I
   answered one question and it didn't remember."

Restarting from the top is NOT the defect and should not be changed here: spec
06 §2.1 says Ctrl-C during onboarding exits without writing anything and the
next run is still a first run, and src/first-run.ts:115-118 implements that
deliberately ("leaving is not answering"). Whether a half-finished onboarding
should ever resume is a separate product decision, not part of this issue —
if it is ever taken, it changes the spec first.

Spec

specs/spec06/06-persona-onboarding.md §2.1 — the input primitive and the
Interruption rule. The contract is satisfied on "writes nothing"; it is silent
on continuing to ask, and on telling the listener where they stand.

Done when

• A /quit (or Ctrl-C) during onboarding stops the question loop at once —
  no further question is put to the listener after the latch fires.
• Leaving mid-onboarding says plainly that the answers were not kept and
  that the next boot starts over, instead of only "stopped before the
  broadcast."
• A unit test covers both against the quit latch: N questions asked before
  the quit, none after, and the closing line present.

Where it is not

Not the first-run trigger. isFirstRun was right to fire here — the run used a
fresh MURMUR_HOME (/tmp/murmur-demo, created 13:03:23); the real home's
~/.murmur/data/memory/persona.md was untouched.