perf(streaming): throttle renderTail DOM updates to rAF cadence

[jdmanring]

Problem

renderTail() in streamingRenderer.js is called on every SSE token. Even with the in-place fast-path patch, every call creates a holder div, sets its innerHTML (full markdown parse → full DOM subtree), reads childNodes, and then discards the holder+subtree if the fast path matched. At 200 tok/sec this is 200 holder divs and 200 full DOM subtrees created and discarded per second in Oilpan.

Capping at requestAnimationFrame cadence (~60fps) reduces these allocations 3× at high token rates with no visible rendering degradation — the display already updates at most once per frame.

Fix

In chat.js, wrap the per-SSE-token _renderStream() call with a rAF guard:

// module-level
let _streamRenderRaf = null;

// in the SSE token handler, replace direct _renderStream():
if (!_streamRenderRaf) {
  _streamRenderRaf = requestAnimationFrame(function () {
    _streamRenderRaf = null;
    _renderStream();
  });
}

The existing StreamRenderer state (lastText) always holds the most recent full text, so the rAF callback renders the latest token even if intermediate tokens were skipped.

Depends on

Issue #XX (renderTail call counter) — fast-path ratio data should be collected first to confirm this optimization is worthwhile.

Classification

Upstream-candidate.

[jdmanring]

Superseded by #64 (fix/dom-oom-streaming-throttle), which implements an identical rAF guard at the same call site. Both were developed independently and are functionally equivalent — same guard logic, same finally-block cleanup. Filing both upstream would produce a duplicate.

[jdmanring]

Reopening: closed as COMPLETED when the fix landed on develop, but per the fork rule an upstream-candidate issue stays OPEN until its upstream PR is filed — and nothing has been filed upstream yet. The staged branch is the active tracker until the PR goes up.