perf: derive deferred-open stand-ins in the renderer instead of shipping them

[lidonggui]

What

On a deferred open (@roobli/md, the default), main no longer attaches a structural stand-in mdast node to every block or ships them over IPC. parseDocument sends nodes: null with nodesEnrichment: 'deferred'. blockSpansFromWire then derives each block's stand-in from its kind and source the first time something reads span.node. Engine-owned blocks never read it; the rest are re-parsed from source when they're enriched.

• lazyNodeSpan(fields, build) returns a span whose node is a memoised getter. The getter is an own, enumerable property, so structuredClone, spreads and toEqual still see a plain value and the BlockSpan contract is unchanged.
• standInNodeFor(kind, markdown) exposes the existing standInNode.
• enrichSpanNone and finalizeEngineAdapterSpan build stand-ins the same lazy way, so main never builds them for the wire either.
• Wire validator: a deferred open may omit nodes. A full open without nodes is still refused. Incremental saves are unchanged (no nodes, no nodesEnrichment).

Measured (huge, 8 MB, 43,970 blocks, Linux box)

	before (f25d1fe)	after
open wire payload (v8 serialised)	24.98 MB	14.07 MB
main publish (IPC send), profiled	~71 ms	~36 ms
deserialise (Node v8 proxy)	~164 ms	~42 ms
main parseDocument, profiled	~310 ms	~330 ms (unchanged within noise)
packaged bench BENCH_RUNS=5, interleaved ×2	1,826 / 1,671 ms	1,636 / 1,582 ms
profiled packaged open, interleaved ×3	1,747 / 1,669 / 1,600 ms	1,890* / 1,516 / 1,499 ms

*outlier

Honest note: building the stand-ins was only ~10 ms of parseDocument, so main parse time doesn't move. Its cost is the engine split plus canSkipDialectEnrich / engineSemanticKey on every block. The win is the payload and IPC: ~100–150 ms of open on huge.

Tests

• pnpm typecheck and pnpm test pass.
• New: derived stand-ins equal the ones main used to ship for every golden fixture (after a structured clone), and build the same ProseMirror document.
• New: lazy stand-ins are built once and only when read.
• New: an incremental-save wire still returns null from blockSpansFromWire.
• Updated the three tests that asserted deferred nodes were non-null.