humanspeak / svelte-markdown / 36570680698

96%

push

github

perf(streaming): bound per-frame work, fix parity bugs, paired Streamdown benchmark (#396)

* perf(bench): measure per-frame streaming work and correct Streamdown claims

The stream-compare benchmark timed `content = slice; await tick()`. Svelte
Markdown defers stream parsing to the next animation frame, so in burst mode
that was ~0.1 ms regardless of document size (all work landed in one settle
frame), and in frame-paced mode it collapsed to the 60 Hz frame period for
both renderers. The "4.5-9x faster under burst" comparison claims derived
from it were measurement artifacts.

- SvelteMarkdown: opt-in User Timing measure `svelte-markdown:stream-flush`
  around the flush (parse + diff + state write), enabled by setting
  `globalThis.__svelteMarkdownProfile = true`; off by default so entries do
  not accumulate in long sessions. Unit tests cover the helper.
- stream-compare page: per-frame main-thread work = sync update time +
  deferred rAF work + forced layout; reports p50/p95/p99/peak, frames over
  16.7 ms, first-vs-last-fifth growth ratio, and the library flush share.
  New corpora mirror LLM output: mixed prose at 1 and 4 updates/frame, one
  long open bullet list, one long open code fence, citation-heavy prose with
  trailing `[n]: url` definitions.
- runner: summarises the new metrics; output-length mismatch is a warning
  flagged in the JSON instead of aborting the suite.
- docs compare-data + competitive-intel config: replace the retired claims
  with the 2026-09-27 measurements against svelte-streamdown 4.2.0. Svelte
  Markdown does ~1.2x less work only when several updates land per frame;
  at one update per frame Streamdown does 1.2-3x less on every corpus, with
  Svelte Markdown dropping frames on citation definitions (one full
  re-render per definition line) and on long open lists.
- evidence (two 5-iteration suites, logs + JSON + README) archived under
  .agents/.plans-closed/stream-bench-frame-work/evidence.

Co-Authored-By: Claude Fa... (continued)

1301 of 1400 branches covered (92.93%)

Branch coverage included in aggregate %.

539 of 541 new or added lines in 9 files covered. (99.63%)

2001 of 2026 relevant lines covered (98.77%)

337644.63 hits per line