perf: defer heavy chat shell renderers

[0xSolarPunk]

Summary

• Defers file viewer/editor surfaces from the default app shell so CodeMirror and markdown viewer chunks are not fetched until a file is opened.
• Defers the rich tool-event renderer until a tool-use message is actually rendered.
• Collapses chat feed render item construction, tool-result lookup, and permission terminal lookup into one pass over visible messages.

Perf notes

• Baseline no-chat shell fetched rgxz0B2U.js at 929,406 bytes and afXDL_8X.js at 711,925 bytes.
• After this change, the no-chat shell no longer fetched the 929 kB CodeMirror chunk; largest initial script observed was D9I-KMdv.js at 650,886 bytes.
• Opening a chat with tool events still rendered correctly; the after chat-open path loaded the deferred tool-renderer chunk without fetching CodeMirror.
• Build output server layout chunk dropped from about 1,322 kB to about 1,177 kB.

Evidence

• Baseline initial screenshot: https://files.catbox.moe/5e95ie.png
• Baseline chat-open screenshot: https://files.catbox.moe/pncseg.png
• After initial screenshot: https://files.catbox.moe/6qnwnx.png
• After chat-open screenshot: https://files.catbox.moe/bhej66.png
• Baseline repro video: https://files.catbox.moe/ri4rn6.webm
• After repro video: https://files.catbox.moe/hybmrq.webm

Dogfood report is at /tmp/garcon-dogfood-perf/report.md on the profiling machine.

Validation

• bun run --cwd web test conversation-feed-items
• bun run check
• bun run test
• timeout 45s bun run start --port 0 --bind-address 127.0.0.1

Startup smoke reached http://0.0.0.0:21369 before the timeout shut down the test server.

Notes

• Compared against the current OpenAI Codex checkout in /tmp/openai-codex; the relevant design takeaway was to avoid unnecessary work in the steady chat render path, similar to Codex TUI's frame diffing approach.
• Checked current Svelte 5 source/docs in /tmp/svelte-official; this keeps the change in canonical runes/await-boundary territory.