Re-arm the readiness deadline after a gate resolves

[nikolasd]

wait_for_readiness's overall deadline was computed once at function entry and never touched
again. The Gate branch correctly parks with no deadline of its own -- a human answering a first-run
gate can take as long as they need -- but the Undecided branch still measured its remaining time
against that same original, never-advanced deadline.

Concretely: a run parks at a gate, a human takes real time to notice and answer it (the entire point
of escalating), the screen advances to something not yet recognized -- and because the original
entry-time deadline had long since elapsed while the run was correctly parked, that very next
Undecided tick failed the run instantly, with a message that reads as though nothing had happened at
all. The human who did exactly what the escalation asked of them was the one punished for it.

No shipped vendor classifier walks this path today: claude and codex's gates both resolve straight to
a screen their own classifiers recognize as the composer, so Gate never hands off to Undecided in
production. Planning a vendor whose first-run flow has more than one screen is what surfaced this
before it shipped.

Fixed with one new bool tracking whether the previous classification was a Gate, checked and cleared
on the next Undecided tick to re-arm the deadline exactly once per Gate-to-Undecided transition --
never on every Undecided tick, which would make fail-closed stop meaning anything for a screen that
is genuinely, permanently unrecognizable after a gate. The per-gate escalation dedup is untouched, a
separate concern. The failure text also now distinguishes the two cases: a gate that was seen and
left gets a message saying so, rather than one implying nothing was ever recognized.

Testing: parking on a gate long enough for the original deadline to elapse, then transitioning to a
briefly-unrecognized screen, must still succeed -- verified as a genuine regression control by
temporarily disabling the re-arm and confirming this test, and only this test, starts failing. A
screen that stays unrecognized forever after a gate must still fail once the fresh window elapses,
bounded well inside an outer timeout, proving the re-arm fires once per transition rather than on
every tick.