feat(chat): custom agent mailbox helpers and session.in delivery fixes (#4644)

## Summary

Adds `chat.messages.hasPending()` and `chat.messages.next()` so a custom
agent
loop can inspect pending chat input without consuming it and take one
record at a
time, and fixes four ways a chat could mishandle input across a restart:
a
message silently lost, a recovered answer cut off by a stop the user had
already
pressed, a retried send answered twice, and a record the agent had no
consumer
for blocking every message queued behind it.

```ts
if (await chat.messages.hasPending()) {
  const record = await chat.messages.next({ timeoutInSeconds: 0 });
  if (record) handle(record.payload);
}
```

## Why the fixes came together

`session.in` carries records for consumers whose delivery needs differ.
A user
message must be delivered eventually, so it can wait arbitrarily long
for a turn
to take it. A stop only means anything to the turn that is live when it
lands.
Progress along the channel was tracked as one sequence number, and one
number
cannot say "control applied through 7, message 3 still owed" at the same
time.
Each of the bugs above is that mismatch surfacing somewhere different.

So instead of a rule per symptom, records are now classified once and
handed to
one route, and each route declares two things: whether it holds a record
when no
consumer is ready, and whether a record it never handled has to survive
into the
next boot. The resume cursor, the replay window and the
discard-the-unowned
behaviour are then derived from route state rather than maintained
beside it, and
`hasPending()` answers from the message queue instead of the head of a
buffer
shared with every other kind.

The wire is unchanged. Both cursors on the turn boundary keep their
meanings, so
existing chats resume as before and there is no webapp change.

## Behaviour worth calling out

`chat.writeTurnComplete()`'s `sessionInEventId` is the cursor that is
safe to
resume from, not the sequence of the record the turn answered. It is
held back
behind any message still waiting to be handled, so a value below the
record you
just handled is expected rather than a sign of a lost turn.

The stop fix also covers chats whose most recent turn was completed by
an older
SDK, by resolving the replay window from the channel when the boundary
does not
carry one. The trade there is deliberate: a stop that landed in the
moments
before boot and was never applied is dropped along with the replayed
ones,
because a stop the user can press again beats a stale one killing an
answer they
are waiting for.

## Verification

Thirteen reproductions against a local stack, each driving real runs
rather than
mocks, covering the documented `next()`/`hasPending()` loop, suspend and
resume, a
crash between consuming a message and writing turn-complete, a retried
send whose
idempotency claim is lost, and a continuation boot that must not replay
answered
messages. Where applicable each was also run against `main`, so the
fixes are
differences rather than assertions.

Five further legs on a deployed environment, which the earlier revisions
of this
branch did not cover at all: a message appended while the run is
genuinely
checkpointed, a message appended while the run is dead, the
stop-after-crash case
on the real crash path, and both version-skew directions (a newer worker
resuming
an older worker's turn boundary, and an older worker resuming a newer
one's).

Two of those restart fixes also have a browser-driven red and green pair
on a
deployed environment, staged identically on both sides and differing
only in the
SDK. For the lost-message fix, the unanswered message is replayed and
answered in
full here, and is never replayed at all on the released SDK. For the
stop fix,
both sides replay the message and diverge on the stop itself: it is
declined here
and the answer completes, while the released SDK re-applies it and the
recovered
answer dies before it streams.

The routing decision itself is a pure state machine, so it also has a
property
test over every interleaving of the record kinds crossed with each crash
point,
checked by mutation to confirm it fails when the cursor arithmetic or
the replay
window is broken.

## Known and not addressed here

The read of the woken record is unbounded, so a wake with nothing to
read makes
`wait()` outlive its own waitpoint. Tested and not a deadlock, since the
read
defers to the next record, but bounding it is a separate change with its
own
test.

Separately, and not caused by this branch: a run that crashes while a
message is
still queued is not replaced until the next inbound append, so that
message waits
rather than being recovered on its own. Worth its own issue.

Also not caused by this branch, but worth knowing when reading the
release note: a
chat page that stayed open across the crash keeps showing the partial
answer it
already received, so the recovered answer only appears after a reload.
The answer
itself is persisted correctly. The gap is on the client, which does not
apply a
re-delivered turn over a partial it already holds.

---------

Co-authored-by: Eric Allam <eric@trigger.dev>
Co-authored-by: Eric Allam <eallam@icloud.com>