Root cause: every sx_server worker session used js-eval on the 3.6KB
HARNESS_STUB, paying ~15s for tokenize+parse+transpile even though every
session does the same thing. Over a full scoreboard with periodic worker
restarts that's minutes of wasted work.
Fix: transpile once per Python process. Spin up a throwaway sx_server,
run (inspect (js-transpile (js-parse (js-tokenize HARNESS_STUB)))), write
the resulting SX source to lib/js/.harness-cache/stub.<fingerprint>.sx and
a stable-name symlink-ish copy stub.sx. Every worker session then does a
single (load .harness-cache/stub.sx) instead of re-running js-eval.
Fingerprint: sha256(HARNESS_STUB + lexer.sx + parser.sx + transpile.sx).
Transpiler edits invalidate the cache automatically. Runs back-to-back
reuse the cache — only the first run after a transpiler change pays the
~15s precompute.
Transpile had to gain a $-to-_js_dollar_ name-mangler: the SX tokenizer
rejects $ in identifiers, which broke round-tripping via inspect. JS
$DONOTEVALUATE → SX _js_dollar_DONOTEVALUATE. Internal JS-on-SX names are
unaffected (none contain $).
Measured: 300-test wide (Math+Number+String @ 100/cat, --per-test-timeout 5):
593.7s → 288.0s, 2.06x speedup. Scoreboard 114→115/300 (38.3%, noise band).
Math 40%, Number 44%, String 30% — same shape as prior.
Baselines: 520/522 unit, 148/148 slice — unchanged.
giles
4a277941b6