[arpan404]

Follow-up to #13, fixing two review findings.

1. Multiline spans still flickered

The hold looked only at the open line. Streaming `hello and then a line break released the opening backtick, and the closing world` took it back when the span became code.

Inline spans now carry across soft line breaks and end with their paragraph, as in CommonMark. The hold starts where an unfinished span opens in the open paragraph. That paragraph starts after a blank line, a fence, a heading, a table row, or a rule or setext underline, or at a new list item or quote.

2. Every delta rescanned the history

shownLength compared the whole previous answer with startsWith and rescanned the open line on every update, so small deltas added up to quadratic work.

• StreamHold.append(source) handles the last source with text added. It processes the lines completed since the last call once each, and the open paragraph from its last settled point.
• StreamHold.replace(source) is the full pass for new text.
• Checks on line shape only look at short lines, so a long open line isn't sliced on every update.
• MarkitRenderer.render decides append vs. replace with one comparison. That decision also picks the worker's append path, replacing the startsWith that #renderPatch made on its own. Callers that know they're appending can pass render(source, { appended: true }) to skip it.
• Finalized renders do no hold work.

Time spent in the hold for 3,000 appends, measured with each string flattened before the call:

history	previous hold	this PR
20 KB	30 ms	~1.4 ms
2 MB	2,263 ms	~1.8 ms

Tests

• tests/node/stream-hold.test.mjs:
  • `hello\n stays held and becomes hello world code once world` arrives;
  • an answer with code, bold and links that cross soft breaks never takes back text, character by character;
  • a paragraph break releases a span that never closed.
• tests/node/stream-hold-cost.test.mjs checks that 3,000 appends cost about the same after 2 MB as after 20 KB (3/3 runs pass), and that a replacement still reads the whole source.
• node --test tests/node/*.test.mjs: 15/15.
• Playwright shared, renderer, inline, code, github, react: 40/40, twice.
• Streaming 257 real agent answers 3 characters at a time: 8 updates take back text, all in one answer about footnotes, the same as Hold unfinished inline markup while a message streams #13.