Incoming JSONRPCNotifications were processed via independent .subscribe()
calls, allowing async handlers to complete out of arrival order. Route
notifications through a Sinks.Many drained by concatMap so each handler
Mono must complete before the next notification is picked up.

Adds testNotificationOrderPreservedWithAsyncHandler to verify that five
rapid notifications with staggered async delays still arrive in order.

closeGracefully() completed the notification sink and disposed the
concatMap subscription in the same synchronous block, so notifications
still queued were discarded. Sync session update consumers are always
async - SyncSpec wraps them with subscribeOn(SYNC_HANDLER_SCHEDULER) -
so a rapid burst was lost on close. Before the sink was introduced each
notification was an independent subscription that outlived the session,
so nothing was dropped.

Await drain termination before teardown, bounded by requestTimeout so a
handler that never completes cannot hang graceful close.

A notification arriving after the sink completes is dropped, which is
correct: a graceful shutdown stops accepting new work while draining what
is queued, and JSON-RPC notifications carry no delivery guarantee. But
that expected case was logged at WARN alongside overflow, zero-subscriber
and non-serialized emission, which lose traffic on a live session.

Log the shutdown cases at DEBUG and the anomalies at ERROR, and name the
method in both.

…acts

The LICENSE file was a paraphrase of Apache-2.0, not the license. It
omitted section 6 (Trademarks) entirely and renumbered 7-9 down to 6-8,
rewrote the section 2 copyright grant, narrowed the "Licensor" and "Work"
definitions, and dropped the appendix. GitHub therefore classified the
repository as NOASSERTION while every published POM declared Apache-2.0.

Replace it with the verbatim text from apache.org, add a NOTICE file, and
publish both under META-INF in every module artifact as Apache-2.0
section 4 expects.

Also drop the redundant <repositories> block: Maven Central is implicit,
enabling snapshots on it is meaningless, and inheriting a repository
definition from a published parent POM is discouraged.

A CycloneDX 1.6 SBOM scan of the published closure reported 17 known
vulnerabilities (5 HIGH, 11 MEDIUM, 1 LOW): 11 in jackson-databind and
jackson-core 2.21.2, 6 in Jetty 12.0.14. Both reach every consumer as
compile-scope transitives of acp-core and acp-websocket-jetty.

Rescanning after the bump reports zero.

Three *IT classes existed but never executed: Surefire's default includes
do not match *IT, and no Failsafe plugin was configured, so `mvn verify`
- which is what CI runs - silently skipped them.

Wire Failsafe. CleanShutdownIT's 4 tests now run and pass; GeminiCliIT's 5
correctly self-skip via @EnabledIfEnvironmentVariable without GEMINI_API_KEY.

All 8 action references used floating tags. release.yml runs with
contents: write and holds the Maven Central credentials and the GPG
signing key, so a retagged or compromised action would execute with
publish rights. Pin each to the commit SHA the tag currently resolves to,
keeping the version in a trailing comment.

The READMEs linked to springaicommunity.mintlify.app, which has retired
this project - its own nav files acp-java-sdk under "Moved to markpollack" -
and which still documents 0.11.0, three releases behind. The maintained
copy is lab.pollack.ai/docs/acp-java-sdk. All six linked routes verified.

Both tests slept a fixed 500ms and then asserted the response had been
sent. The handler runs asynchronously, so on a loaded machine the sleep
expires first: the test failed 4 of 5 local runs at the previous commit
and twice in a row on CI.

Poll for the expected message to a 10s deadline instead. 10 consecutive
local runs pass, and the happy path is faster than the old fixed sleep.

Correctness and supply-chain hygiene: notification ordering and the
graceful-close drain, Jackson/Jetty CVE remediation, the verbatim
Apache-2.0 license and packaged LICENSE/NOTICE, and the build fixes.
No protocol or public API changes - verified: one main-source file
changed since v0.14.0, with no public or protected signature changes.

No job declared timeout-minutes, so a wedged runner would hold the job
for the 6-hour default before giving up - which is exactly what happened
twice on main, at a step that completes in about a minute locally.