perf(mac): incremental streaming Markdown — parser path now costs the same as the fast path at real scale

[SergeSerb2]

Summary

Closes the streaming-perf P0 that #450 measured: any answer containing a bullet, link, or code span fell onto an O(n²) main-actor reparse (pinned p95 1,166 ms, only 224/1,819 deltas consumed in the 60 s soak). Four compounding fixes, each measured:

1. List-safe split boundaries: the scanner now promotes boundaries before zero-indent list markers — long streaming lists (the most common assistant shape) previously never settled any prefix. The cache renumbers a lazily numbered ("1./1./1.") ordered run continuing across a seam so streamed display equals the full parse; nested/indented markers still refuse the split. Every-split equivalence with the exact parser is property-tested (fixtures cut at every character offset).
2. Frozen-prefix projection: ParsedMarkdownDocument.settledBlockCount lets both the fragment projection and the store's render-row replacement reuse the settled prefix verbatim — per-soak render-row projections fell 3.65 M → 10 k.
3. Validated append hint reaches the cache: no more O(accumulated) exact prefix compare per delta (was 24.6 s of the soak's main-actor time; the exactness fix from fix(mac): AppKit markdown emphasis/appearance and sidebar selection fixes #441 stays as the fallback for anything without a valid hint).
4. Content-gap recovery: a pure content-lane revision gap applies the slice's absolute item values instead of a full presentation rebuild. The rebuild used to outlast two stream ticks and cascade — nine rebuild-gap-rebuild chains per soak, now zero. Structural/metadata/environment gaps still rebuild; append hints are ignored on the recovery path.

Render-row diffs now trust the engine's revision contract instead of comparing AttributedStrings per fragment; MarkdownMessageChrome drops the source string (copy actions read the semantic row) so it stays O(1)-comparable.

Measured (new structured-prose payload + scenes — the shape providers actually send; commands + artifacts in the parity matrix):

• 150 KB bounded: pinned p95 23.6 ms vs the plain fast path's 23.3 ms in the same harness (pre-fix: 87.7 / 219.9 / 286.1 ms) — parser path == fast path at realistic scale.
• Adversarial 60 s / 3.7 MB / ~12,000-block single message: all 1,817 deltas consumed (was 224), zero fallbacks, pinned p95 ~170–192 ms. Residual cost is the update protocol's O(rows) ID arrays per mutation — recorded in the matrix as the follow-up; it only bites at that scale.

Area

• apps/mac — native macOS app
• apps/windows
• apps/mobile
• apps/server
• Shared packages or relay
• Build, CI, or release tooling
• Docs

Release size

• size:XS
• size:S
• size:M
• size:L
• size:XL

Verification

• pnpm run verify --all after merging current main (incl. fix(mac): optimistic send — the message appears when you press Enter #451/feat: narrate worktree setup — echo first, typed phase progress, setup script on the eager path #452): 5/5 steps, exit 0
• New tests: list-boundary promotion + nested refusal, lazy-list seam renumbering, bounded-tail assertion (200-item list), every-split equivalence fixtures (lazy + loose lists), gap-recovery covered by existing suites staying green
• Perf scenes run 5×, evidence under .artifacts/incremental-md-*, honest numbers in the parity matrix

🤖 Generated with Claude Code