[Mihir-Rabari]

Root cause

SplitViewState.clearAllPanels reset the split tree by assigning a fresh SplitNode.Panel("main") to _rootNode and nothing else. The outgoing root's BossTabsComponents were simply dereferenced: their tabs' LifecycleRegistrys stayed RESUMED, so onDestroy never fired - and for browser and terminal tabs those handlers are what release the Chromium process and the shell process. Since WorkspaceApplier.applyWorkspace calls clearAllPanels() between every apply (#1310), each restore over an unpreserved tree leaked one native process per open tab, compounding across session restores and workspace applies.

Design

The outgoing tree is disposed before the fresh main panel is assigned, using the exact primitive closeCurrentWorkspace already uses - clearAllTabs() on every panel - so per-tab lifecycle destruction (removeTab(i, recordForOpen = false) -> tabLifecycles.remove(it.id)?.destroy()) finally fires for the dropped tree.

One guard makes this safe: disposal is skipped when the outgoing root is still identity-held by preservedWorkspaceStates. A tree kept by preserveCurrentState must survive to be restored on switch-back, and its live tabs keep moving through moveTabToWorkspace and collectAllActiveTabs while preserved - destroying them would break the cross-workspace switch that SplitViewWorkspaceTransferTest and SplitViewActiveTabsTest pin. The identity check (it.rootNode === outgoingRoot) distinguishes the two cases without touching any preserved-state semantics.

getAllPanels() is read before _rootNode is reassigned, so it walks exactly the outgoing tree.

Tests

New SplitViewClearAllPanelsDisposalTest with a destroy-counting tab component (lifecycle onDestroy -> counter):

• clearAllPanels destroys outgoing tree tabs when no preserved state holds it - both tabs of the old root destroyed exactly once; fresh main panel starts empty.
• clearAllPanels keeps a preserved outgoing tree alive for switch-back - a tree recorded by preserveCurrentState keeps its tab undestroyed and reachable through panelsInWorkspace.
• clearAllPanels disposes every panel of a split outgoing tree - disposal covers both panes of a vertical split.

Gates

One gradlew --offline ktlintCheck :composeApp:detekt :composeApp:desktopTest --tests 'ai.rever.boss.components.window_panel.*' --max-workers=2 invocation:

tests="3" skipped="0" failures="0" errors="0"   SplitViewClearAllPanelsDisposalTest (new)

Full ai.rever.boss.components.window_panel.* filter green - 222 tests completed, 0 failures (all pre-existing suites unchanged, including the preserved-workspace and transfer pins).

Fixes #1310

[Mihir-Rabari]

CI note: the only red on this PR is a pre-existing macos timing flake — expected to clear on a failed-job re-run.

• Failed job: build-test (macos-latest) (run 35499290639, job 106048095974, 08:35–08:51Z). build-summary is only its aggregate. Ubuntu and Windows on this same SHA are green.
• The failure is TerminalLimitsTest > cancellation during return dispatch terminates the unclaimed process — IllegalStateException: Expected a dispatched continuation raised by PausedDispatcher.next(), a hard queued.poll(5, TimeUnit.SECONDS) wait (TerminalLimitsTest.kt:55, driven from line 231) for a continuation from a real process-spawning session.
• This PR's diff is confined to composeApp (window_panel: SplitViewState.clearAllPanels + SplitViewClearAllPanelsDisposalTest). It does not touch the boss-app-terminal module, so it cannot affect that test's dispatch timing.
• The same suite is green on build-test (macos-latest) for fix(updater): require word-boundary flanks for inline emphasis in release notes #1228/fix(network): capture retryAttempt before the Checking write so offline auto-retry counts attempts #1229/fix(workspaces): restore nested-split tabs exactly once and keep restorable siblings behind unknown tabs #1319/fix(window): retire pending initial tabs and projects when a window closes #1320/fix(kernel): require verified caller identity on the five unattributed service bridges #1339/fix(launchpad): cap session-preserved staging paths per plugin #1342 at the identical base (5dda86e) — only this run hit the 5s deadline, on a job that itself took 15m28s in the currently loaded runner pool.

Everything else: 8/10 checks green, MERGEABLE. A re-run of the failed job should clear it (PR-author tokens cannot re-run repo workflows).

[Mihir-Rabari]

Rebuilt on current dev (bc9aa3b) as an isolated 2-file change (SplitView.kt, SplitViewClearAllPanelsDisposalTest.kt); the branch no longer carries unrelated history. Fresh CI is starting on 8cbad70.

[Mihir-Rabari]

Fresh CI on 8cbad70: build-test (macOS/Ubuntu/Windows), code quality, compatibility, database, contracts, version, and shell checks all passed. Ready for re-review.

[fluck-boss]

Merged to dev (squash). Gate: full CI green on the merge preview (10/10 checks) + maintainer diff verification - clearAllPanels now disposes the outgoing tree's tabs before reassigning _rootNode, releasing browser/terminal native processes (#1310), and the exemption for a tree still held by preserveCurrentState is explicit so switch-back keeps working. The new test pins both the disposal and the preserved-tree survival. No follow-ups filed. Thanks @Mihir-Rabari!