- Skip recording event history (snapshots of all frontend models + the
  backend model per event) when tests run via toTest, toSnapshots,
  startHeadless, or the viewer's run-all pass, since history is only
  read by the interactive test viewer. This makes memory use flat
  instead of growing linearly with simulated time.
- Skip storing pending effects when a command is Command.none, and
  check pending effects with Array.isEmpty instead of folding over
  every entry each simulation step.
- Find the earliest timer end time with a direct fold instead of
  allocating intermediate lists on every animation frame step.

Benchmark (1 simulated hour, 100ms frontend timer + 1s backend timer +
600 network roundtrips): heap 67MB -> 9MB, wall time 1.92s -> 1.71s.
Test pass/fail results and failure messages are unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XHzxkUaYwwvczLWDasAcL

perf-harness/run.py copies an app (e.g. at-chat, meetdown), generates a
headless startHeadless runner for its Effect.Test suite, compiles it once
against a baseline program-test and once against a new version (both
injected into an isolated ELM_HOME as the version the app pins), runs both
in node, verifies the test output is byte-identical, and reports wall
time / max RSS / final heap medians.

run-tests.js includes an XMLHttpRequest shim so addStringFile and
addBytesFile viewer file loads are served from the app directory.
seed-elm-home.py pre-fills an ELM_HOME package cache from a donor cache
plus jsDelivr, for sandboxes where GitHub archive downloads are blocked.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XHzxkUaYwwvczLWDasAcL