In the course of poking at this, I noticed that https://h.readthedocs.io/projects/client/en/latest/developers/developing.html 's section on testing is fairly egregiously out-of-date. We should update that documentation after we sort this.

After doing some digging, it looks like the issue is that the --watch flag matches a V8 flag which is detected by the flagged-respawn package that Gulp uses internally. According to the npm README, flagged-respawn respects a --no-respawning flag to disable its behavior. This almost works:

gulp test --watch --no-respawning

Except that the runTests function in @hypothesis/frontend-build complains about not understanding the --no-respawning flag. With a slight modification to the CLI arg parsing, the above command works:

  // Parse command-line options for test execution.
  program
    .option(
      '--grep <pattern>',
      'Run only tests where filename matches a regex pattern'
    )
    .option('--watch', 'Continuously run tests (default: false)', false)
    .parse(process.argv.filter(a => a !== '--no-respawning'));

This is why various other tools that have a --watch flag continue to work in the latest version of Node: they don't use gulp. Normally the Node-specific (as opposed to tool-specific) --watch flag would only be handled when the flag comes before the script name in the command:

node --watch path/to/tool --watch

In the above the first --watch is for Node, the second --watch is for the tool.