perf(streaming): intra-list tail bounding — freeze settled items of an open list (fixes #29)

[jonathanKingston]

Fixes #29 — the biggest perf gap left after #23: a long, still-open trailing list kept its whole group in the tail (loose/tight is a whole-list property, so no item could settle), re-rendering/re-sanitizing/re-morphing every item on every commit. That's the old O(n²) for arguably the most common long LLM output shape.

Result: streaming a 200-item list drops 8,870ms → 1,053ms (8.4×); growth per doubling ~4.1× → ~2.1×.

How

Settled items freeze inside the shared <ul>/<ol> (two sibling lists would break byte-parity — CommonMark renders one), per the design on #29:

• render-blocks.ts — the list-group knowledge is factored into exported primitives (listGroupSignature, listSliceContinuesGroup, listItemSliceIsMultiParagraph, scanListGroup, renderListItemsSlice, listGroupOpenTag/CloseTag) and collectListGroup is rewired onto them: rendering and streaming share one source of truth for group membership, looseness, and item HTML. An item slice renders byte-identically to the corresponding items of a whole-group render — the property everything else rests on.
• streaming-dom-morph.ts — morphElementChildrenFrom (reconcile an element's children from an offset) and syncAttributes (the shared list element may legitimately change attributes — e.g. gaining class="contains-task-list" when a later item adds a checkbox — while its frozen <li>s are never touched).
• streaming-frozen-tail.ts — the intra-list mode. Activation is pure state init after a generic commit whose trailing group is a signature-uniform open list with ≥4 items starting at the frozen boundary. Per commit, the unfrozen item slice is scanned (incremental loose evidence, exactly scanListGroup's rule), all but the last item freeze (the last can still absorb continuations; token extension over the boundary trips the existing straddle guard), and the region is reconciled with one wrapped-list morph so settling items keep DOM node identity. Sealing: when a non-continuing token ends the group, the list is promoted to a fully frozen top-level node and the generic path handles what follows in the same commit — no full-morph at list end.

Safety (tight-optimistic, everything behind the existing full-morph fallback)

• Tight→loose flip against frozen items → one-time fallback per list (looseness is monotonic, so amortized O(n); the rebuilt state freezes loose thereafter).
• Signature break, unbalanced benign raw-inline tags in a would-freeze slice, in-place rewrites (frozenSource byte guard), DOM shape mismatch → fallback.
• Mixed-marker giant runs and long open blockquotes stay on the generic path (conservative; noted in the module header — blockquote grouping re-renders merged content and isn't sliceable the same way).

Verification

• 472 tests + CommonMark conformance + exhaustive convergence fuzz green.
• Full-history byte-parity (every committed frame, char-by-char) through freeze → seal → following ordered list → seal, and each hazard: tight→loose flip, late task-class, marker change, late multi-paragraph item, start attribute, nested lists, rewrite-while-active.
• Frozen-<li> node identity preserved for the rest of the stream.
• Deterministic CI guard for the list shape: renderedChars grows 1.95–1.99× per item-count doubling (< 3× threshold; ~4× = regression).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LnAFvAgDY3RVYT6zJqN1fy

---

Generated by Claude Code