ServerKey's keyed <Show> remount is a multi-second synchronous cascade (dispose + rebuild of the whole app subtree) that used to leave the UI looking frozen. A tiny module-level serverSwitching signal now gates a fullscreen Splash rendered above the ServerKey boundary, and the status-popover click handler setTimeout-defers the batched navigate+setActive so the browser paints the splash before the freeze begins and dismisses it after the new subtree paints.
Serialize non-Error promise rejections so unhandled rejections print type/ctor/keys/JSON instead of the unreadable '[object Object]'. Also emit [server health] logs when a health poll returns unhealthy and when polling switches servers, so a red dot in the status popover comes with a logged URL and auth presence. Minor cosmetic: restore session-header StatusPopover import position after the earlier titlebar experiment.
Terminal onCleanup ran persistTerminal synchronously during a dispose cascade, which flowed through props.onCleanup -> ops.update -> update() in context/terminal.tsx and fired setStore on the terminal store. That store write reentered the reactive graph mid cleanNode iteration; solid then nulled an ancestors owned while an outer cleanNode recursion was still iterating it, crashing with Cannot read properties of null reading 1 at node.owned[i]. Wrapping finalize in queueMicrotask pushes the store write past the current synchronous cleanup cascade so the teardown cannot race with cleanNodes owned walk.
Hooks DEV.hooks.afterCreateOwner to wrap every owner's .owned and .cleanups with accessor traps that record every mutation to a ring buffer with tags, stacks, and cleanup-depth context. On any 'Cannot read properties of null' TypeError the buffer is dumped so the offending cleanup/origin that nulled an owner's owned mid-iteration is visible post-hoc. Also wraps owned arrays in a Proxy so cleanNode's index reads are logged and the suspect ownerTag at crash time can be identified. Debug only; zero cost until a crash fires.
Same nested-dispose-in-onCleanup bug as 7f36ac2481 but in three more
places: TerminalProvider.disposeAll, PromptProvider.disposeAll, and
scoped-cache.clear() (covers viewCache.clear and comments cache.clear).
All of them synchronously call createRoot dispose() on cached entries
inside onCleanup, which during a server switch nests into the outer
cleanNode cascade and throws TypeError at chunk-*.js:992.
Snapshot the pending disposers, clear the cache synchronously, and
fire the disposers on a microtask so the outer cleanup finishes first.
disposeDirectory called a createRoot dispose() synchronously. When
triggered by pinForOwner's onCleanup during a parent remount (e.g.
switching to a WSL server re-keys the ServerKey Show), the inner
dispose ran a nested cleanNode cascade on a sibling root while the
outer cascade was mid-traversal, corrupting solid-js's graph walk
state and surfacing as TypeError: Cannot read properties of null
(reading '1') at chunk-*.js:992 after ~155 recursive cleanNode frames.
Queue the dispose on a microtask so synchronous bookkeeping still
runs (map deletes, onDispose cache invalidation) but the reactive
cleanup happens after the outer traversal finishes.
The status popover and select-server dialog used to call navigate('/') then
defer server.setActive to the next microtask. With multiple sidecars in v2,
that split triggered two separate disposal cascades - one for the route
change and a second for the ServerKey Show re-key - and the sidebar project
bucket also swaps (local -> wsl:Debian), tearing down every solid-dnd
sortable in the middle. Wrapping both calls in batch() lands them in a
single Solid update so disposal runs once.
Also raise Error.stackTraceLimit to 200 so future disposal crashes capture
the originating frame instead of truncating at the tenth cleanNode.
Electron's console-message event only surfaces {level, message, line, sourceId}
without the stack, so uncaught errors showed up as 'line 1028 of chunk-*.js'
(SolidJS's rethrow site) with no way to find the real origin. Attach
window error and unhandledrejection listeners that log the full stack via
console.error, and reshape the main-process log line so newlines in the
stack survive instead of being JSON-escaped into one unreadable blob.
The main process was resetting webContents zoom to 1 on every
\zoom-changed\ event, which fires not just for native Chromium zoom
gestures but also for the renderer's own setZoomFactor IPC calls. Paired
with a keydown listener that re-sent the current zoom on every
ctrl-combination (ctrl+backspace, ctrl+z, ctrl+v, ...), this created a
self-triggered race that intermittently snapped the factor back to 1.
Make the renderer the single source of zoom truth: keyboard, wheel, and
menu all drive the same Solid signal, preventDefault blocks Chromium's
built-in accelerators before they race, and the main process disables
pinch zoom and no longer listens to zoom-changed.
Local Server is always Windows-native now; WSL lives as a separate list
of one-or-more distro-bound sidecars spawned alongside it. Manage Servers
shows an Add WSL button on Windows, each WSL server appears as its own
row with remove + retry, and the wizard runs scoped to a new distro.
