<!-- Generated by community/.ai/render-guides.mjs; edit community/.agents/skills/fix-project-leak-from-tc-report/SKILL.md -->

Fix a _LastInSuiteTest.testProjectLeak from a TeamCity report

This is the full fix workflow. Investigation-only tasks (identify + reproduce + analyze) are covered by the sibling leaking-test-investigation-by-tc-report skill; the two overlap in phases 1–3 but this one adds phases 4–5 (apply fix, verify).

The workflow is:

1. Identify the culprit test from the TC report.
2. Reproduce the leak locally 10 times, archiving every heap dump + stdout.
3. Analyze all unique retention chains observed (local ± CI-reported).
4. Fix each retention chain independently with a pattern from the catalogue below.
5. Verify by rerunning the leak-hunter loop on the patched tree and confirming it stops firing.

The bundled scripts are self-contained — do not source the other skill's copies.

Phase 1 — identify the culprit test

The leaked Project name is the strongest signal. In ProjectRule() / heavy-project rules the project name defaults to the test class simple name. So:

• Look for Instance: Project(name=<TestClassSimpleName>, containerState=DISPOSE_COMPLETED, …) in the TC report.
• Locate the file with ijproxy search_symbol / search_file; only ask the user if the class name matches multiple files and the retention chain does not disambiguate.
• Record the test module (.iml next to the test's testSrc/) — this is the --module for tests.cmd.

If the project name is generic (light_temp_…), pick the first non-platform class from the retention chain (ProjectCodeStyleSettingsManager, a plugin service, a Mockito mock, …) and search callers/tests referencing it plus the plugin area.

Fallback ordering (in decreasing signal strength): heap Instance line → CI hashcode branch → retention-chain class names → the user.

Phase 2 — reproduce locally 10 times

_LastInSuiteTest.testProjectLeak is filtered out of pattern discovery (community/platform/testFramework/core/src/com/intellij/TestCaseLoader.java:181-182,382-384); its body runs via JUnit5TestSessionListener.testPlanExecutionFinished (community/platform/testFramework/src/com/intellij/tests/JUnit5TestSessionListener.java:39-193), gated on UsefulTestCase.IS_UNDER_TEAMCITY == true, i.e. env TEAMCITY_VERSION != null.

To activate locally without pretending to be a real agent, set TEAMCITY_VERSION and point TEAMCITY_BUILD_PROPERTIES_FILE at a two-line stub properties file (needed so TeamCityHelper.getPersistentCachePath in the wrapper doesn't crash):

mkdir -p /tmp/tc-stub-cache /tmp/tc-stub-tmp
cat > /tmp/tc-stub.properties <<'EOF'
agent.persistent.cache=/tmp/tc-stub-cache
teamcity.build.tempDir=/tmp/tc-stub-tmp
EOF

Then run the bundled loop script — do not hand-roll 10 sequential tests.cmd calls in the harness:

"${CLAUDE_SKILL_DIR}/scripts/leak-loop.sh" \
  --module <test-module-iml-name> \
  --test  <test-FQN> \
  --runs  10 \
  --archive /tmp/<name>-leak-runs

leak-loop.sh sets the two env vars for you (auto-creating the stub properties file if missing), loops N times, copies each run's Heap dump is published to <path> file to $ARCHIVE/run-NN.hprof.zip, saves stdout to run-NN.output, and prints one boundary line per run so a Monitor can subscribe.

Fast-exit contract. If run 1 produces no heap dump (no leak line in the output), the script aborts with exit code 3 and does not queue the remaining runs. Treat this signal explicitly:

Exit code	Meaning	Interpretation during Phase 2
0	All N runs finished; all N produced a leak.	Deterministic leak. Proceed to analysis.
3	Run 1 produced no heap dump; loop aborted.	Either the leak is not reproducible locally, or the leak-hunter didn't fire. Debug before continuing.

If exit code is 3 in Phase 2, verify that the run's stdout contains an ideaTests.totalTimeMs line (comes from JUnit5TestSessionListener.testStatistics after the leak check) — if yes, testProjectLeak fired and found no leak (unreliable repro; check the test suite membership, ordering, and any co-located classes). If no, the listener didn't fire — recheck TEAMCITY_VERSION propagation.

Cold Bazel compile + JBR download makes run 1 take 4–5 minutes; subsequent runs are ~30–60 s each with warm caches.

Phase 3 — analyze every unique retention chain

After Phase 2, invoke:

"${CLAUDE_SKILL_DIR}/scripts/summarize-runs.sh" /tmp/<name>-leak-runs

It writes SUMMARY.md with a per-run table plus the set of unique retention-chain shapes (class-only, argument values stripped). This tells you whether all runs leaked the same way or you have multiple parallel bugs.

Then hand-write <archive-dir>/ANALYSIS.md — required before writing any fix — covering:

1. What reproduced. Test FQN + module, N/N run ratio, list of unique retention chains.
2. Each unique retention chain, both local and the one reported by CI (if different). Print the full via '<field>' path with class names, and label the chain (e.g. Chain L / Chain C).
3. Root cause per chain. Locate the retaining code via search_symbol and cite file:line-range.
4. Recommended fix per chain — pick from the catalogue below.

Do not skip the CI-reported chain even if you only reproduced a different one locally. The leak checker reports one path per project; multiple retainers can coexist on the same leaked instance. Fixing only one usually promotes the next to be reported next time.

Phase 4 — apply fixes

MANDATORY first check — is the retaining class a Disposable project service?

Before you apply any pattern from this catalogue, answer this question about the class that ends up as a Disposer ROOT (or otherwise pins the Project):

Is the class both

1. registered via <projectService> in a plugin descriptor or annotated @Service(Level.PROJECT) (equivalently @Service with a project-level scope inferred by the container), and
2. a Disposable (implements Disposable directly or via a base class)?

If YES, then the service container itself already registers this instance into the Disposer tree and disposes it correctly. Specifically:

• community/platform/service-container/src/com/intellij/serviceContainer/ServiceInstanceInitializer.kt:61-62 — on construction, the container calls Disposer.register(componentManager.serviceParentDisposable, instance). The service is parented under serviceParentDisposable, which is itself a child of the project (or app) in the Disposer tree. The service is not supposed to be a Disposer ROOT.
• community/platform/service-container/src/com/intellij/serviceContainer/ComponentManagerImpl.kt:904-905 — on disposal, the container calls Disposer.dispose(instance) (not instance.dispose()), which cascades to every child registered under the service, invokes instance.dispose(), and removes the instance's entry from ObjectTree.myObject2ParentNode.

Consequences for the fix:

• Disposer.register(this, child) inside such a service's constructor (typical uses: new Alarm(SWING_THREAD, this), messageBus.connect(this), EP.addExtensionPointListener(project, listener, this)) is the canonical, correct pattern. Do not change this → project (or any other parentDisposable) at those call sites. That reduces Disposer tree height but does not fix any leak — see "Anti-pattern — flattening the Disposer tree by hoisting children to project" below.
• If the retention chain still terminates at ObjectTree.myObject2ParentNode with the service as a Disposer ROOT, then something on the framework path above did not happen for this specific case. The actual bug lives there. Investigate, in this order:
  • Constructor-vs-container-registration ordering / re-registration semantics. The service's constructor runs Disposer.register(this, alarm) before the container runs Disposer.register(serviceParentDisposable, this). If the second call is a no-op / rejected because this is already present in ObjectTree with an implicit ROOT parent, the service stays a ROOT even after the container "registers" it. Read ObjectTree.register / Disposer.register for the reparenting rules.
  • Light-project reset path. LightPlatformTestCase.tearDown and TestApplicationManager release reuse projects across test classes and have custom reset semantics. If they don't route through the container's Disposer.dispose(instance) cascade, the ObjectTree entry lingers. Recent tracking commits are under IJPL-247543 — start there with git log --grep IJPL-247543 -- '**/serviceContainer/**' '**/LightPlatform**' '**/LeakHunter**'.
  • The leak-hunter's timing relative to Disposer.dispose(project) cascades. A race window can report a ROOT that would have been cleaned up milliseconds later.
  • Recent framework changes. Grep the same paths under IJPL-247543 for context.

If NO (the class is not a Disposable project service, or it is manually instantiated / registered outside the standard service container path), fall through to the "Root cause first" section and pick a pattern.

Root cause first — fix disposal, not the reference

A Project reaches the leak-hunter because some object holds it. That object falls into one of three categories, and the correct fix depends on which:

• (A) Legitimately outlives the Project by design — an application-scope cache, message bus, extension point, thread-local, static field. Then the retention itself is the bug: break it (unregister the listener, clear the cache on projectClosed, add a parentDisposable), or — only when unregistration is genuinely impossible — weaken the reference at the retention boundary.
• (B1) Should NOT outlive the Project AND is a properly-registered Disposable project service — registered via <projectService> / @Service(Level.PROJECT), implements Disposable. The framework path (ServiceInstanceInitializer.kt:61-62 + ComponentManagerImpl.kt:904-905) is expected to dispose the instance and remove it from ObjectTree. If the retention chain shows the service still a Disposer ROOT, something on the framework path did not happen for this case — investigate the possibilities listed in the "MANDATORY first check" above. Do not paper over it by moving children out from under this.
• (B2) Should NOT outlive the Project AND is NOT integrated with the service container — a manually-instantiated helper, a project-scoped listener registered elsewhere, a project-scoped CoroutineScope, or anything registered under the Project's Disposer subtree by test / production code that bypasses the container. Here the retention is a symptom of a structural disposal bug and a real fix is warranted: register it properly (Disposer.register(project, this) in the constructor, migrate to @Service(Level.PROJECT), route disposal via Disposer.dispose(...) at the call site, add a missing parentDisposable, unregister the listener, evict from the cache on projectClosed).

(B1) is by far the more common case for a service that appears as a Disposer ROOT, and neither WeakReference on myProject nor flattening the child-Disposer registrations is a valid fix for it. Wrapping a service's own myProject field in WeakReference only makes the Project GC-able while leaving the underlying disposal bug in place — see the anti-pattern below. Moving children out from this to project reduces Disposer tree height without fixing why the service itself lingers in ObjectTree — see the other anti-pattern below.

Only reach for WeakReference after you have identified case (A) and confirmed the container legitimately outlives its contents.

Anti-pattern — WeakReference on a service's own Project field

Do not "fix" a project-service leak by wrapping the service's own myProject field in a WeakReference. Example of what NOT to do:

// BAD — hides the disposal bug, does not fix it.
private final WeakReference<Project> myProjectRef;
public MyService(Project project) { myProjectRef = new WeakReference<>(project); }
private Project project() { return Objects.requireNonNull(myProjectRef.get(), "disposed"); }

Why this is wrong:

• A project service must be disposed together with the project. If the retention chain shows a project service pinning its own Project, the service is outliving the project — that is the actual bug. The most common concrete culprits (for cases outside category (B1) — see the MANDATORY first check for the (B1) framework path first):
  • A service registered a listener on an application-scope extension point / message bus without a parentDisposable tied to the project.
  • A service registered itself into a static / application-scope cache and no projectClosed handler evicts it.
  • A non-standard registration path bypassed ServiceInstanceInitializer so the container never routed the service through Disposer.dispose(instance).
• Weakening the field only masks the Project reference. The service instance itself is still leaked (still reachable from wherever the disposal bug is). Message-bus subscriptions still fire on the "dead" service; Disposable children of the service stay in the tree; log lines and telemetry keep referencing the disposed project by name.
• Objects.requireNonNull(myProjectRef.get(), ...) starts throwing NPE from unexpected paths (isDisposed() checks, teardown callbacks, log formatters) as soon as GC clears the reference — usually intermittently and only under load.
• Every future GC pause, heap-shape change, or class-loader tweak can silently re-expose the leak by making the retention checker report the SERVICE (still leaked) or a downstream field instead of the Project. You have not fixed the leak; you have moved the leak checker's crosshair. TC will keep reporting a leak; the report just points somewhere else.

Correct approach for a service-in-the-Disposer-tree chain — diagnosis flow:

1. Verify container integration. Confirm the retaining class is registered via <projectService> in a plugin descriptor or annotated @Service(Level.PROJECT), and implements Disposable. If YES, the framework path from community/platform/service-container/src/com/intellij/serviceContainer/ServiceInstanceInitializer.kt:61-62 (registration under serviceParentDisposable) plus community/platform/service-container/src/com/intellij/serviceContainer/ComponentManagerImpl.kt:904-905 (Disposer.dispose(instance) on shutdown) should dispose the service and remove its ObjectTree entry. If the chain still shows the service as a Disposer ROOT, the bug is on the framework path or in the disposal ordering, not in the service's constructor. Investigate why cleanup didn't happen for this specific case (see the "MANDATORY first check" possibilities: constructor-vs-container ordering / re-registration semantics, light-project reset path, non-standard registration, leak-hunter timing, recent IJPL-247543 commits). In this case Disposer.register(this, child) in the constructor is canonical and must not be changed.
2. Only if the service is NOT properly integrated with the container (manually instantiated by test / production code, custom registration path that bypasses ServiceInstanceInitializer, or a non-service Disposable that ended up as a Disposer ROOT), apply a structural registration fix. Options, in order:
   • Migrate the class to @Service(Level.PROJECT) / <projectService> so the container owns its Disposer wiring.
   • Register the instance into the project's Disposer subtree explicitly at the construction site with Disposer.register(project, instance).
   • Route the disposal call through Disposer.dispose(service) where the caller currently invokes service.dispose() directly (that leaves the ObjectTree entry behind).
   • Add a missing parentDisposable at the listener / EP registration site.
3. Never as the first move: flattening children into project. Rewriting Disposer.register(this, child) to Disposer.register(project, child) inside a Disposable service's constructor only reduces Disposer tree height — it does not remove the service from ObjectTree. See "Anti-pattern — flattening the Disposer tree by hoisting children to project" below.
4. Only after these are exhausted and you have confirmed with a repro loop that the service genuinely cannot be disposed in time (an infrastructure bug that is out of scope for the ticket), consider surface-level containment — and prefer a test-side dispose in tearDown or an explicit unregistration hook over field-weakening.

Anti-pattern — flattening the Disposer tree by hoisting children to project

Do not "fix" a project-service leak by rewriting Disposer.register(this, child) in the service's constructor to Disposer.register(project, child) (or to any other higher-level parentDisposable). Example of what NOT to do:

// BAD — reduces Disposer tree height but does not fix the leak.
public DbPsiFacadeImpl(@NotNull Project project) {
  Disposer.register(project, new Alarm(SWING_THREAD, this)); // was: Disposer.register(this, ...)
  Disposer.register(project, project.getMessageBus().connect()); // was: connect(this)
  EP.addExtensionPointListener(project, listener, project); // was: parentDisposable = this
}

Symptom this often gets confused with: the retention chain terminates at ObjectTree.myObject2ParentNode with the service instance itself as a Disposer ROOT (e.g. ObjectNode.myObject → MyServiceImpl@..., ObjectTree.myObject2ParentNode → ..., (root) → ObjectTree, Disposable chain to Disposer ROOT: MyServiceImpl@... <- ROOT).

Why it's tempting: removing the Disposer.register(this, ...) calls in the constructor appears to "remove the ROOT node" (children no longer keep the service pinned via their parent-node backref) and a single local run of the leak-hunter can look clean.

Why it's wrong:

• For a properly-registered Disposable project service (<projectService> / @Service(Level.PROJECT)), children under this are already disposed via the container's Disposer.dispose(instance) cascade — see community/platform/service-container/src/com/intellij/serviceContainer/ComponentManagerImpl.kt:904-905. Disposer.register(this, alarm) is the canonical way to tie the child's lifetime to the service's lifetime and it is not the source of the leak.
• Moving those children to project bypasses the service's own disposal cascade and gives the children the project's lifetime instead of the service's lifetime. If the service is later replaced (dynamic plugin unload, replaceServiceInstance in a test, or any hot-reload path), the old service's children linger under project until project close — a fresh leak / behavioural bug you introduced.
• It does not fix the retention chain in CI. The container is still responsible for cleaning up the service's own ObjectTree entry (see ServiceInstanceInitializer.kt:61-62 and ComponentManagerImpl.kt:904-905); whatever prevented that from happening still prevents it after your edit. The next TC run will report the same service or the next-most-downstream retainer.

Correct alternative: treat the retention as a bug on the framework's disposal path, not in the service's constructor. Investigate ObjectTree.register / Disposer.register reparenting semantics, the light-project reset path (LightPlatformTestCase.tearDown, TestApplicationManager), non-standard registration, and the leak-hunter's timing (see the "MANDATORY first check" and IJPL-247543). Preserve Disposer.register(this, child) in the service constructor.

Chain families

Two chain families cover the vast majority of testProjectLeak failures. Match your observed chain to one of the patterns:

Pattern A — Mockito thread-local retention on the EDT

Signature: chain terminates in MockingProgressImpl.ongoingStubbing → Thread[AWT-EventQueue-…].threadLocals → IdeEventQueue (root). The retained argument array contains the leaked Project.

Root cause. Every mock invocation goes through Mockito's MockHandlerImpl.handle(...), which writes the current Invocation into a MockingProgress ThreadLocal and never clears it. Tests annotated @RunsInEdt run on the EDT, so this ThreadLocal accumulates on the EDT and outlives the test class.

Preferred fix. Replace mock<Foo>() + whenever(...) with a small hand-rolled fake. Example (from ApkEditorTest.kt, before/after):

// BEFORE
private fun mockFileEditorProviderManager(): FileEditorProviderManager {
  val mgr = mock<FileEditorProviderManager>()
  val provider = mock<FileEditorProvider>()
  whenever(provider.accept(any(), any())).thenReturn(true)
  whenever(provider.createEditor(any(), any())).thenAnswer { TestFileEditor(it.arguments[1] as VirtualFile) }
  whenever(mgr.getProviderList(any(), any())).thenReturn(listOf(provider))
  return mgr
}

// AFTER
private class FakeFileEditorProvider : FileEditorProvider {
  override fun accept(project: Project, file: VirtualFile): Boolean = true
  override fun createEditor(project: Project, file: VirtualFile): FileEditor = TestFileEditor(file)
  override fun getEditorTypeId(): String = "test-fake"
  override fun getPolicy(): FileEditorPolicy = FileEditorPolicy.NONE
}
private class FakeFileEditorProviderManager : FileEditorProviderManager {
  private val providers = listOf(FakeFileEditorProvider())
  override fun getProviderList(project: Project, file: VirtualFile) = providers
  override suspend fun getProvidersAsync(project: Project, file: VirtualFile) = providers
  override suspend fun getDumbUnawareProviders(project: Project, file: VirtualFile, excludeIds: Set<String>) =
    providers.filterNot { excludeIds.contains(it.editorTypeId) }
  override fun getProvider(editorTypeId: String) = providers.firstOrNull { it.editorTypeId == editorTypeId }
}

Fallback fix (when a fake is too much work — many abstract methods, awkward suspend signatures, etc.): hoist the mocks to class fields and reset them on the EDT after each test.

private val mgr = mock<FileEditorProviderManager>()
private val provider = mock<FileEditorProvider>()

@After
fun clearMockitoEdtState() {
  runInEdtAndWait {
    Mockito.reset(mgr, provider)          // clears MockingProgress ThreadLocal on this thread
  }
}

The Mockito.reset(...) route resolves the ThreadLocal on whichever thread it's called on, so the runInEdtAndWait { } wrapper is load-bearing.

Pattern B — Extension-point / message-bus listener with a strong this capture

Signature: chain contains ExtensionPointImpl.listeners (or MessageBusImpl.subscribers) → <listenerAdapter>.handle → a lambda / anonymous inner class whose arg$1 is the retaining service, whose field points at the leaked Project.

This pattern applies at a retention boundary where an application-scope container (ExtensionPointImpl.listeners, MessageBusImpl.subscribers) legitimately outlives a project-scope service by design. That is a category (A) case — the app-scope container is supposed to live longer than any single project. WeakReference here weakens the reference at the boundary between scopes, on the lambda's captured this. It does not weaken the service's own myProject field (see anti-pattern above).

If your chain does not go through ExtensionPointImpl.listeners / MessageBusImpl.subscribers and instead terminates in ObjectTree.myObject2ParentNode or a project-scope holder, this is not Pattern B — do not apply this fix. Go back to the "Root cause first" section.

Root cause. A project-scoped service registered a listener on an application-scoped extension point using addChangeListener(listener, parentDisposable). ExtensionPointImpl.addChangeListener adds the adapter to listeners immediately and only removes it in a Disposer.register(parentDisposable) { removeExtensionPointListener(adapter) } cleanup. If that cleanup does not run (or runs after the leak check), the adapter — and everything the listener lambda captures, including the service and its Project field — stays alive at the application scope.

Preferred fix. First, verify that removing / unregistering the listener at project dispose is genuinely not feasible (usually because the parentDisposable disposal path is not under your control). Only then, break the strong this capture by wrapping in a WeakReference at the lambda boundary. Example (from CodeStyleSettingsManager.java:199-208, before/after):

// BEFORE
protected void registerExtensionPointListeners(@Nullable Disposable disposable) {
  FileIndentOptionsProvider.EP_NAME.addChangeListener(this::notifyCodeStyleSettingsChanged, disposable);
  ...
}

// AFTER
protected void registerExtensionPointListeners(@Nullable Disposable disposable) {
  // WeakReference indirection: the app-scoped extension point outlives project-scoped subclasses,
  // so a missed disposal here would pin the manager (and its Project) via the listener list.
  java.lang.ref.WeakReference<CodeStyleSettingsManager> selfRef = new java.lang.ref.WeakReference<>(this);
  FileIndentOptionsProvider.EP_NAME.addChangeListener(() -> {
    CodeStyleSettingsManager self = selfRef.get();
    if (self != null) self.notifyCodeStyleSettingsChanged();
  }, disposable);
  ...
}

Why this is safe: while the service container holds a strong reference to the manager, selfRef.get() returns non-null and the listener behaves identically. After container release (project dispose for a project service), the WeakReference can clear and the extension point's listener list no longer pins the manager. Same trick applies to anonymous-inner-class listeners — pull the enclosing this into a WeakReference local and use ref.get() inside every callback method.

Do not simply remove addChangeListener. The listener has a real behavioural purpose; you must preserve that when the service is alive.

Other patterns (rarer)

• Static/application cache holding a project reference. Retention chain terminates in a static field or an application service's Map/List. Fix: use WeakHashMap / ContainerUtil.createWeakMap keyed by project, or explicitly clear on ProjectManagerListener.projectClosed.
• Background coroutine/executor holding a captured project. Retention chain terminates in a CoroutineScope.job node or a Thread.target runnable field. Fix: scope the coroutine to project.coroutineScope (auto-cancelled on project close) instead of GlobalScope / Dispatchers.Default.
• MessageBus subscription without a disposable. Retention chain contains MessageBusImpl.subscribers. Fix: use the subscribe(Topic, listener, parentDisposable) overload, not the two-arg one.

Phase 5 — verify

Rerun the same loop, targeting the same archive directory (or a fresh one). The loop's fast-exit is the desired positive signal here:

"${CLAUDE_SKILL_DIR}/scripts/leak-loop.sh" \
  --module <test-module-iml-name> \
  --test  <test-FQN> \
  --runs  3 \
  --archive /tmp/<name>-leak-verify

Interpretation of the verify run:

Signal	Meaning
Exit code 3 + first-run ideaTests.totalTimeMs line present + no Found a leaked instance.	Success: testProjectLeak ran and passed.
Exit code 0 + N/N runs still producing hprofs.	Fix did not close the leak. Reopen ANALYSIS.md and look for a chain you missed (or a wrong pattern).
Exit code 3 + no ideaTests.totalTimeMs in the stdout.	The leak-hunter didn't fire. Recheck TEAMCITY_VERSION propagation and the properties stub.

For extra confidence do 2–3 additional plain tests.cmd invocations after the loop's success, greping the output:

export TEAMCITY_VERSION=local-leak-check TEAMCITY_BUILD_PROPERTIES_FILE=/tmp/tc-stub.properties
for i in 1 2 3; do
  ./tests.cmd --module <mod> --test <FQN> > /tmp/verify-$i.out 2>&1
  leak=$(./community/tools/rg.cmd -c 'Found a leaked instance' /tmp/verify-$i.out || echo 0)
  stats=$(./community/tools/rg.cmd -c 'ideaTests.totalTimeMs'   /tmp/verify-$i.out || echo 0)
  echo "run $i: leaked=$leak listener_ran=$stats"
done

Expected: every run prints leaked=0 listener_ran=1.

Reporting

Finish with a compact table like:

	Before fix	After fix
Found a leaked instance occurrences	N/N runs	0/M runs
Heap dump is published occurrences	N/N runs	0/M runs
Listener ran (ideaTests.totalTimeMs)	N/N	M/M
Test JVM exit code	41 (leak)	0 (clean)

Attach the diff of the fixes (test-side + platform-side) and the path to the archived SUMMARY.md / ANALYSIS.md. This is the deliverable — enough for a PR reviewer to reproduce, understand, and sign off.

Guardrails

• WeakReference is not the default fix. The default fix for a project-scope retention is to fix disposal — remove a misplaced Disposer.register(this, child) only when the enclosing class is NOT a properly-registered Disposable project service, add a missing parentDisposable, unregister the listener, evict from the cache on projectClosed. WeakReference only applies at a genuine cross-scope boundary (Pattern B). Wrapping a service's own myProject in WeakReference is an anti-pattern that hides the disposal bug rather than fixing it (see Phase 4 "Anti-pattern"). If you find yourself reaching for WeakReference<Project> on a service field, stop and re-read the "Root cause first" preamble.
• For a Disposable project service, Disposer.register(this, child) is canonical — do NOT remove it. The service container itself parents Disposable services under serviceParentDisposable (community/platform/service-container/src/com/intellij/serviceContainer/ServiceInstanceInitializer.kt:61-62) and disposes them via Disposer.dispose(instance) (community/platform/service-container/src/com/intellij/serviceContainer/ComponentManagerImpl.kt:904-905), which cascades to all children registered under this. Rewriting Disposer.register(this, child) → Disposer.register(project, child) in a Disposable service's constructor only reduces Disposer tree height and does NOT fix any leak (see Phase 4 "Anti-pattern — flattening the Disposer tree by hoisting children to project"). If a Disposable project service still appears as a Disposer ROOT in the retention chain, the bug is on the framework path — investigate the "MANDATORY first check" possibilities before touching the constructor.
• Never edit _LastInSuiteTest, JUnit5TestSessionListener, or TestCaseLoader to work around a leak — those are the leak checker itself. If the check is spuriously failing, that is a testFramework bug, not a per-test fix.
• Do not silence the leak checker locally (e.g. -Dintellij.build.test.ignoreFirstAndLastTests=true) as a "fix" — that only hides it.
• Do not shorten a WeakReference wrapper into a soft/phantom reference. Soft references defer collection under memory pressure and will not close the leak deterministically. Phantom references cannot be get()-ed.
• Do not commit changes to /tmp/tc-stub*.properties — those are throwaway files, not repo state.

Key source references

• Culprit-test filter for _LastInSuiteTest: community/platform/testFramework/core/src/com/intellij/TestCaseLoader.java:181-182,382-384.
• Suite hook that fires testProjectLeak: community/platform/testFramework/src/com/intellij/tests/JUnit5TestSessionListener.java:39-193.
• TC-gate flag: community/platform/testFramework/src/com/intellij/testFramework/UsefulTestCase.java:143.
• Leak checker: TestApplicationManager.testProjectLeak() — search TestApplicationManager.kt under community/platform/testFramework/src/com/intellij/testFramework/.
• Extension-point listener disposal contract: community/platform/extensions/src/com/intellij/openapi/extensions/impl/ExtensionPointImpl.kt:730-735.
• tests.cmd semantics: .claude/skills/testing/SKILL.md.