Every merge builds the catalogue, and a step then checks the output. One check makes sure each inline <script> parses, because a page whose script has a missing semicolon renders fine and does nothing. It did this by starting node --check once per distinct script body: about 1,460 processes per run.
A CPU profile showed the time was in spawning and waiting on those children, not in parsing. One child now compiles all of the bodies in-process (nothing runs) and reports the ones that parse. A body it does not vouch for still goes to node --check, so the pass/fail verdict and its message still come from node. A local run of the step went from 28.2 s to 17.3 s, with identical output and 97 passing tests. The CI figure will come from the next pool reading.
Details: docs/history/ci-perf-2026-11-09-inline-parse-batch.md.