perf(frontend): break down the startup bundle and bound frontend leaks

[Harry19081]

Summary

Frontend RAM pass, first batch: break down the startup bundle, fix the real cause of the duplicated-vendor chunks, and bound the frontend growth paths that were leaking — all behavior-preserving except the terminal mount window (see risks). Ranked audit with measured numbers lives in docs/memory-audit-2026-08-16/ram-optimization-findings.md.

Measured with webpack --mode production at HEAD before/after:

	Before	After
Synchronous initial JS (index.html scripts)	8.37 MB (main 4.74 + vendors 3.73)	0.98 MB (main 0.32 + vendors 0.65)
JS loaded at boot (initial + App import set)	≈ 8.4 MB	≈ 5.9 MB (−29 %)
Vendors loaded at boot	3.73 MB	≈ 1.65 MB
Modules emitted in >1 chunk	1 601 / 9.0 MB redundant	339 / 1.9 MB
Total build/	55 MB	48 MB

Problem

• src/index.tsx chose between the dev-only webpackMode: "eager" App import and the normal one via an isDev const. Webpack can't fold that, walks both arms, and "eager" wins → production shipped App (and every vendor only App needs) inlined into main.js as ~8 MB of synchronous startup JS.
• The chat-projection web worker statically imported the registry barrel, which re-exports rendering/registry/events/index.ts and its lazy renderer import()s. That made 2 388 files part of the worker's chunk graph; since the worker entry has no vendors, webpack duplicated react-dom / xterm / CodeMirror / zod / sql-formatter into every shared async chunk (~9 MB).
• Four sidebar files, SidebarDialogs, ModalSystem, ActionSystemContext and the boot-registered editor actions imported barrels that dragged xterm + addons, all of CodeMirror + 10 language packs, sql-formatter, refractor (for a modal nobody renders) and framer-motion into the boot graph.
• Several frontend structures grew without bound: unsavedContentCache (two full copies of file text per edited file, never evicted), runtimeCounters.durations[] (one entry per RPC, no drain call site), search "load more" accumulation, per-session todo slots, output channels (+ a sessionStorage rewrite of every channel on every appended line), and every initialized terminal staying mounted forever (incl. all restored at boot).

Solution

• Inline process.env.NODE_ENV === "development" at the App import branch.
• Move CONTEXT_CONFIG + chat-config helpers to registry/events/contextConfig.ts (pure data); ActionRegistry/registryAccessors import from it. Worker reach with async edges 2 388 → 62 files. splitChunks unchanged.
• Deep imports at the boot gateways; Message toast renderer split into a React.lazy MessageContainer (API unchanged); EditorService loaded on first editor-action invocation.
• Bounds: unsavedContentCache drops the never-read originalContent, tracks dirty, evicts only clean entries past 32 (dirty = truly unsaved edits are never evicted); runtimeCounters running sum/count; searchAppendResultsAtom enforces the documented 20 000-match ceiling; sessionTodoMapAtom LRU 64 (never the active session; rebuilt from events by useTodoSync); useOutputChannels capped at 16 channels with a debounced sessionStorage mirror.
• TerminalCore keeps the active pane + 4 most-recently-active mounted (terminalMountWindow.ts); colder panes unmount and reattach through the existing attach_pty_stream + serialized-buffer restore path.
• Regression guards: src/test/staticImportGraph.ts (fast static import walker) backs startupGraph.test.ts (heavy packages / feature trees stay out of the boot graph, eager import stays foldable) and workerStaticGraph.test.ts (worker never reaches the renderer loaders or react-dom again).

Potential risks

• Terminal mount window (only behavior change): switching to a terminal that has been cold for 5+ activations remounts xterm — brief blank, scroll position/selection reset, and if output arrived while it was cold the restore is the Rust-side bounded snapshot rather than full scrollback. Browser-tab webviews were deliberately not touched (a prototype was reverted; documented in the audit).
• First toast loads its chunk lazily (local file, sub-ms in Tauri; auto-dismiss timers start on mount so no shortened lifetime).
• Editor actions (Go to line / Find / Undo …) do a one-time dynamic import of EditorService on first use.
• The static-graph guards are regex-based (not a full resolver); they'll fail loudly with the import chain if a barrel re-export sneaks a heavy lib back in — bump the allow-list with a reason if that's ever intended.

Validation / Test plan

• pnpm typecheck clean; eslint clean on all touched files (pre-commit lint-staged passed).
• pnpm test: full suite green before the last test additions (1104 files / 8696 tests); new suites green: startup/worker graph guards, Message lazy container, cache/todo/search/output-channel bounds, terminal mount window, contextConfig boundary, lazy EditorService.
• Production builds before/after (table above); module-id duplication measured with an acorn walk over build/.
• Manual: open >5 terminals, cycle through them (reattach + restore), trigger toasts, Go-to-line/Find/Undo from the palette, sidebar rename/share dialogs, quit-confirmation modal.