[morisil]

Fixes #141

Stacked on #161 (fix of #143, itself stacked on #158, #153 and #148) — the base of this PR is issue-143-close-cancels-await, so the diff shows only this fix.
Merge the PRs below it first; GitHub then retargets this PR to main.

Fix

Both races between close() and execute() are closed:

• closed is an AtomicBoolean, and close() uses compareAndSet, so only one call cancels, queues the disposal and shuts the thread down — a concurrent second close() has no effect, instead of throwing RejectedExecutionException;
• execute checks closed again on the session thread, inside withContext, before touching the compiler — so a snippet passing it was queued before the disposal, which always remains the last task on the thread;
• asCoroutineDispatcher() is replaced by a small dispatcher which hands a rejected block to Dispatchers.IO without cancelling it, where its only effect is to fail that check with IllegalStateException("The session is closed")
  — before, a rejected snippet failed with CancellationException: The task was rejected, as if the caller was cancelled, and one queued after the disposal with IncorrectOperationException: … has already been disposed.

The KDoc of execute (@throws IllegalStateException, also for a session closed while the snippet waits for the ones running concurrently) and close() describe it.

DefaultScriptSession takes a new internal executor constructor parameter, a test seam to make the interleavings deterministic; the public ScriptSession(…) factory does not expose it, and the ABI dump is unchanged.

Tests

• should reject snippet of session closed before dispatching it to the thread of the session — closes the session from within the dispatch, so it is rejected;
• should reject snippet of session closed before dispatching it to the thread of the session after the disposal — the snippet, past the early check, is queued behind the disposal;
  both deterministic, failing before the fix (the reviewer re-verified it by removing the on-thread check);
• should close session concurrently from many threads — a stress guard only: the race window is a single volatile read, so it did not fail before the fix either (see Harden the closing of a session: deterministic test, dispatcher fallback #162).

Review

One round of /code-review high, no blocking findings: the ordering of the CAS, the disposal and the shutdown was verified, and the custom dispatcher keeps what kotlinx.coroutines expects (isDispatchNeeded, the resumption on the caller's dispatcher).

Follow-ups filed

• Harden the closing of a session: deterministic test, dispatcher fallback #162 — a deterministic test of concurrent close(), the Dispatchers.IO fallback catching (hypothetical) resumptions, the dispatcher's toString().

🤖 Generated with Claude Code

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.