fix: stream the body progressively past an open last node, and reject instead of hanging on stream errors

[aralroca]

What

Two walker fixes plus a test hardening pass for arbitrary chunk boundaries. Version bumped to 0.6.8 (`deno.json` was stale at 0.6.6 and is now synced).

Fix 1 — the body was never applied progressively when content hangs off the last child

In the `nextSibling` hop, `isLastNodeOfChunk` fell through to `return streamInProgress` — an unconditional wait — even when the open node already had complete children to diff. Since real pages hang their content off the last child of a container (`

` after ``), the walk stalled before entering it and nothing of the body was applied until the stream closed; only the `` (which precedes it) streamed. Measured on a real docs site: during a 3s mid-stream pause, 43 mutations — all in ``, zero in the body.

The fix descends into an open last node once it has children (same condition the `firstChild` hop already used), and defers the pruning phase of `setChildNodes` for a level whose parent is still open — trailing nodes may yet arrive in a later chunk, so removing them early would corrupt the page. Deep clones settle before cloning so a half-parsed subtree is never photographed.

Fix 2 — an erroring stream hung `diff()` forever

`processStream()` only cleared `streamInProgress` on the `done` path. A stream that errors (an aborted fetch, a dropped connection) left it `true`, so the walker waited on `requestAnimationFrame` indefinitely: `diff()` neither resolved nor rejected, plus unhandled rejections from the fire-and-forget reader and `pipeTo`. For a SPA router this meant one aborted navigation froze all navigation until a hard reload.

Now the error is captured, `streamInProgress` clears in `finally`, the duplicate `pipeTo` rejection is silenced, and `diff()` rejects with the stream's error in bounded time.

Tests

TDD throughout — every fix landed with a failing test first. New coverage (ran on Chromium, Firefox and WebKit; 137 pass):

• tag names, attribute names and attribute values split across chunks
• multi-byte UTF-8 characters split at byte level across chunks
• `<script>` semantics pinned: a NEW streamed inline script is inserted but not executed (parsed in the inert document, the "already started" flag survives cloning — identical in all three engines); a pre-existing script is not re-executed
• stream error before/after the first chunk → `diff()` rejects
• progressive body: `h1` inside ``-as-last-child updates mid-stream; trailing siblings added or removed by a later chunk reconcile correctly once the level closes

Harness additions: `newHTMLByteChunks` (raw byte chunks), `errorStreamMessage`, `midStreamEval` (DOM snapshot during the pause before the last chunk).

Why now

The Janux framework moved to streamed SSR (flushing many small chunks with boundaries anywhere) and its SPA navigation aborts superseded page streams — both fixes are prerequisites for that to work, and the chunk-boundary tests pin the contract it relies on.

🤖 Generated with Claude Code