fix(resync): publish off the reconcile pass's thread - #437

Problem

A reconcile pass could not complete while resync publication was making no progress.

ResyncSupervisor::install_live and deactivate block on an acknowledgement from the resync
worker thread, once per live seat per pass. That same thread ran every publication, and a
publication is neither cheap nor bounded: emit_admitted takes the shared catalog-authoring lock,
re-resolves the whole catalog through discover_strict, and takes the recipient's stream lock. So
a publication that was slow, refused-and-retried, or blocked on another process serialized the
whole pass behind it.

That is the property which turned a noisy retry into a two-hour outage in #431. A resync refused
because its recipient is suspended is re-armed every 500 ms, and every attempt re-resolves the
catalog. The refusals had power only because they denied the reconcile pass that would have ended
them — the first completed pass tears the suspended seats' tasks down, and a refusal needs a live
task.

The coupling is observable directly. Holding the recipient's resources/streams/resync/.lock —
the exact flock the publication path takes — and then running a pass:

before   the pass had not returned after 20s; it completed only once the lock was released
after    the pass completes while the publication is still blocked

Goal

A reconcile pass completes whether or not resync publication is making progress, so no publication
condition can deny the pass that would resolve it.

Decisions

• Publication moved to its own thread, rather than bounding the acknowledgement with a
  timeout. A timeout would leave the coupling in place and hide it behind a deadline; the
  acknowledgement exists to order watch installation against launches, which is a read, not a
  publication.
• The worker stays the single writer of carrier state. The emitter reports outcomes back
  through the worker's own mailbox (Msg::Emitted) instead of touching baselines or deadlines.
• One publication outstanding per subscription (Entry.in_flight). A reserved event identity
  must not be handed off twice while the outcome that decides whether it is spent has not returned.
• Queued publications are cancelled when a subscription is deactivated or removed by a
  refresh, so nothing is published to a seat the pass has already decided receives no events. A
  publication the emitter already started is left to finish: it was in flight while the
  subscription was still active, and its outcome is discarded.
• One re-arm mechanism. Re-observing the carrier after an outcome covers both a retained
  refusal and a carrier that moved on while the publication was outstanding. The first draft also
  set dirty and a deadline in the refusal branch; that was redundant, and while it was present
  it masked the re-observation clause under mutation (removing poll_paths was caught by no
  test). With the redundancy gone, that clause is caught by three.
• The emitter is not joined at teardown. A publication blocked on another process's lock must
  not hold up supervisor teardown, and the emitter owns no state to hand back.
• Terminal-refusal classification is deliberately not in this PR. It is still wanted — it
  removes the CPU burn and the unbounded retry — but on its own it would leave this coupling
  intact for the next condition that starves pass completion.

Verification

Fail-before / pass-after — run::tests::reconcile_pass_completes_while_a_resync_publication_is_blocked
holds the recipient's stream lock, confirms the publication has not completed, then runs a pass. A
watchdog releases the lock only if the pass fails to finish on its own, and the test asserts the
watchdog did not have to:

before   FAILED in 22.73s — "the pass only completed after the blocked publication was released"
after    ok — 7 passed (this test plus the six existing resync pass tests) in 2.57s

Mutation — each shipped clause removed one at a time, full resync + resync-pass test selection
run each time. No survivors:

The harness is self-checking: a run that does not compile, or that produces no test-result line, is
reported as UNRAN rather than as a surviving mutant. That guard earned itself — an earlier run of
the same matrix reported every mutant as surviving because cargo was not on the PATH of the
process that invoked it.

Gate — the exact test selection the Nix check runs (--lib --bins --test discovery --test codex_hooks --test hooks --test run --test driver_expansion, --test-threads=1) is green locally:
every target ok, lib 680 passed / 0 failed.

Suite — cargo test -p st2 --lib: 680 passed, 0 failed. --test resync,
--test resync_notify_chain, --test event_e2e, --test reconcile, --test doctor: all green.
Every existing resync unit test still asserts exactly what it asserted before; they now drive the
worker and emitter halves in the order the loop runs them, through a #[cfg(test)] helper.
Production has one path.

docs/vrs/06-resync/spec.md gains the property this depends on: where publication runs, one
outstanding publication per subscription, and what happens to a queued or already-started
publication when a subscription goes away.

Complexity

One thread and one bounded queue. The queue exists because the two operations have genuinely
different liveness requirements: watch installation must answer a reconcile pass promptly, and
publication may block on another process. Merging them is what produced the defect. The alternative
— a mutex over carrier state released around each publication — needs the same outstanding/outcome
bookkeeping without the thread separation, so it is not simpler.

Concerns

• Resync delivery is still serial. A publication blocked indefinitely now stalls resync
  delivery for every recipient behind it, though no longer any reconcile pass. This PR strictly
  improves on the previous behaviour, where the same condition also stalled the passes; refusal
  classification removes the case that made it likely.
• A started publication can land just after a deactivation acknowledgement. Previously
  publication was synchronous on the worker thread, so ordering was total. The window is now
  bounded to a publication that was already in flight while the subscription was active. The
  existing invariant test for that boundary
  (dead_resync_seat_is_deactivated_before_its_relaunch_blocks) still passes.
• The reconcile-pass stall of Terminal resync refusal is retried at 500 ms forever, re-resolving the catalog under the shared authoring lock and starving all writers #431 is not claimed to be fully explained by this change. That
  question stays open there, and the pass rate should be measured separately rather than assumed
  restored.

Friction & bottlenecks

• INVARIANTS.md names three proofs in tests/resource_profile_supervisor_e2e.rs that no longer
  exist, so tests/invariants.rs::qualified_proof_references_resolve fails. It fails identically
  on pristine main (verified against origin/main's INVARIANTS.md), and it is outside the Nix
  gate's test selection, which is why it has stayed red. The correct repair needs the author's
  knowledge of which tests now prove that invariant, so it is not guessed at here.

• tests/agent_publish.rs fails 9 of 22 with host 'host' must declare exactly one root agent; found 0. Reproduced with this branch's files reverted to the merge-base content, so it is
  pre-existing and unrelated. Also outside the Nix gate.

• The devshell's Nix wrapper prints error (ignored): opening log file ... trace.jsonl: Permission denied on every invocation. Harmless, but it is the first line of every build log.

• Overlaps PR Wake the supervisor promptly after catalog publication #433 on INVARIANTS.md and src/run.rs. Different regions — that PR adds a
  catalog-wakeup invariant row and touches src/watch.rs; this one appends a row and adds a test
  inside src/run.rs's existing test module. Whoever merges second resolves a trivial
  append-vs-append conflict in the table.

Follow-ups

• Terminal-refusal classification for Terminal resync refusal is retried at 500 ms forever, re-resolving the catalog under the shared authoring lock and starving all writers #431 (suspended recipient keeps its pending transition and
  gets no retry deadline; the three permanently terminal refusals are dropped with one
  diagnostic). Separate PR, same lane.
• Authoring-time rejection of an over-long PTY session socket path (PTY session socket paths exceeding the sun_path limit are admitted at authoring time, so the seat can never spawn and fails every reconcile pass #432). Separate PR, same lane.
• The two pre-existing red test lanes above.

References

Refs #431