perf(renderer): subscribe panes per session and add memo boundaries instead of re-rendering the whole tree per frame

[Juliusolsson05]

Symptom

useWorkspace() (src/renderer/src/workspace/hook/index.ts:100) selects the entire workspaceRuntimes map and returns a fresh unmemoised object every render (hook/index.ts:897-1021 → App.tsx:87 → WorkspaceContext.tsx:11-17). Every setRuntimes (app-state/workspace/slice.ts:48-51, new map identity per screen/semantic/jsonl message) re-executes App → MainSurface → GlobalEditorShell → TileTree → every TileLeaf / AgentTerminalLeaf / TerminalLeaf / PaneHeader (none memo), including hidden tiled tabs and the display:none retained tree; PaneHeader receives the entire map (TileLeaf.tsx:795). Only Feed (Feed.tsx:294, memo) bails. Live capture: the two renderers burn 55–75% of a core against main's 13–35%; renderer long tasks p99 7.4 s.

Three effects also run over ALL sessions on every replacement: useFeedDebugPersist (paced by #750), usePickerSanity (hook/invalidation/effects.ts:126-138), the codex transcript observation outbox (lifecycle/codexTranscriptObservationOutbox.ts:184-236); getRuntime returns a new identity per call (helpers.ts:87-92).

Expected

• Panes subscribe to state.workspaceRuntimes[sessionId] with useShallow (the pattern already used in DispatchAgentList.tsx:298-305); memo on TileTree, TileLeaf, AgentTerminalLeaf, TerminalLeaf, PaneHeader; narrow PaneHeader's selector; split the workspace context into stable actions and state slices.
• The three all-session effects move to per-session subscriptions or a 1 s timer.
• Migrate consumer by consumer (~25 consumers rely on the fresh object today).

Impact

Re-renders per frame N panes → 1; estimated renderer CPU −30–50% in multi-pane workspaces. Prerequisite for the screen-interest work to pay off fully.
Refs #103, #390. Source: architecture audit 2026-09-03 (scratchpad/arch-perf-audit.md).

[Juliusolsson05]

Follow-up audit inventoried 692 production TS/TSX files across desktop, provider renderers and remote client. PR #785 removes shared-store no-op notifications and unchanged settings serialization (#781), plus closed modal scans (#782); PR #783 reduces terminal transport-triggered acknowledgement and memoizes the scalar-prop inline host. The broad context/per-pane migration remains open: 22 surface wrappers consume workspace context, whose methods explicitly close over current snapshots. No blanket comparator/memo change was made. CommandPalette already correctly gates its expensive inner component except pending invocation; Feed already has memo/index caches. Theme-specific invalidation is now separately tracked in #784. No percentage FPS improvement is claimed from this source audit.

[Juliusolsson05]

Implementation starting in a dedicated perf/pane-runtime-isolation branch from current main 5d64184. First remove runtime/draft subscription invalidation from App, give pane ownership boundaries per-session subscriptions, and preserve reactive Reader/header/debug consumers plus synchronous action/ref freshness and outbox ordering. Existing uncommitted experimental worktrees are reference material only and will remain untouched. Regression tests will count unrelated-pane/layout renders and verify fresh actions, session switching, lifecycle observers and draft persistence. No estimated CPU percentage will be claimed without measurements. Worktree reconciliation is independently implemented in #808; screen interest #762 and corrected diagnostic scope #767 follow separately. No overlap with A6 #802-805.

[Juliusolsson05]

Implementation increment #810 is now open. A real controller/store/tile regression compares the previous root subscription with the new boundaries: 100 updates to one session produce 100 controller/unrelated-pane renders in the control and zero in the isolated path; affected pane stays at 100. Six regressions cover fresh drafts, persistence, focus/session replacement, hidden lease cleanup and lifecycle ordering. Final typecheck passed. Earlier full renderer run passed; final run had 501 passes and one unchanged lazy-prose timeout tracked by #700, documented in the PR. CI is running. Broad inspection consumers remain intentionally reactive; no merge performed.

[Juliusolsson05]

Verification follow-up for #810: final Node 24 renderer run with two workers passed all 115 files / 502 tests in 24.39 seconds. The prior Node 25 lazy-prose timeout (#700) is retained in the PR record; no test timeout or assertion was changed. Full typecheck passed; CI quality-gate is still running.