perf: send an opened document across IPC once

[lidonggui]

What

The openPath, openDialog and activateTab replies no longer carry the document. They name the note ({ path, documentId }), and the document reaches the renderer once, on documentOpened. Every open and activation already publishes that event before it replies.

Why

A packaged renderer CPU profile of opening the 8 MB bench document showed the preload spending ~490 ms in isNotoDocumentWire, almost all of it in its plain-JS SHA-256 of the text. That check ran twice per open: once for the documentOpened event (which App.tsx adopts) and once for the invoke reply, which echoed the same FileTruthOpenReplyV1 but was only read for ok. Main also serialised the ~24 MB payload twice.

• WorkspaceOpenReplyV1.opened is now WorkspaceOpenedNoteV1 | null. A null still means a dismissed dialog or a tab that is not open.
• session.openPath returns an internal WorkspaceSessionOpenV1 (main keeps the full reply per tab). The narrowing happens at the IPC boundary in open-reply.ts.
• The reply validator refuses a reply that carries the document again, so an echo can't creep back in.

Measured

Packaged app, Linux x64 under Xvfb, BENCH_RUNS=5 BENCH_IDLE_LAG=0, same box, before (main 33db02c) and after:

doc	open before	open after
large 2 MB	797 ms	674 ms
huge 8 MB	2,226 ms	1,694 ms

The Linux baseline matches the macOS CI run on e6d373c (huge 2,137 ms). Profiled openPath to first paint: 2,264 → ~1,645 ms.

Verification

• pnpm typecheck and pnpm test pass (2,231 tests), including the new workspace-open-reply.test.ts.
• Packaged e2e on Linux: tabs, shell (open from the empty state, recent list, outline), quick-tabs, reopen, session-restore, code-viewer, links, trail and external-change. 42/42 pass.

Next residual (huge, from the same profiles)

main parseDocument ~320 ms (about half is building stand-in mdast that engine-owned blocks never read), the one remaining preload hash ~200 ms, docFromSpans ~140 ms, EditorState.create ~110 ms (typoraMarks fullScan ~50 ms), and createEnrichFlags re-running canSkipDialectEnrich ~50 ms.