max / quasi

Run the shipped scripts in a browser, and fix the two bugs that found Decided in quasicoherent 4e7f8380 2026-08-30, replacing a deno decision made the same day and withdrawn on re-measurement. Twelve .js files under src/ reach every described document and not one line of them was executed by a test. The Rust tests assert that the script tag is in the head, that without_htmx drops it, and that the constant and the default path agree -- all three true of a file that does nothing. THE HARNESS. mocha and chai as script tags, then the script under test as shipped, then the assertions. This is what Big Sky does for htmx itself, and copying it is not incidental: these are htmx handlers, and htmx exists to exploit real browser behaviour. No npm, no package.json, no node_modules, no bundler, and not a line changed in the twelve files to make them testable -- each is a bare IIFE that installs its own listeners, which is the right shape for a <script> and the wrong shape for every import-and-call runner. run.py is standard library only. WebDriver is REST over HTTP, so a runner needs no package; a requirements.txt beside a suite whose argument is that it needs no package manager would be answering itself. Firefox rather than Chrome because Google publishes no Linux arm64 chrome-for-testing and astra is aarch64, so one browser is one story rather than two. AND IT FOUND TWO BUGS, which is the argument for it. d131a14c, the fold. `settle` started its walk at `null` and only ever learned about a *descendant's* chevron, never the pressed row's own state, so shutting a branch hid nothing at all. A reader could open a branch and not close one. Seeded from the row that was clicked. The second is new, and the emitter's own comment already claimed otherwise ("filters by it, pressing its chevron does not"). The chevron listener bubbled, so its stopPropagation stopped nothing that had already run: a listener on the row is below the document in the bubble path and fires first. node.rs hangs a table row's activate on the row element and filters it with `click[!event.target.closest('[data-act]')]`, which a chevron is not -- so pressing the chevron on a row that both discloses and activates folded the branch and navigated away from it. The listener captures now, which is the one place that knows what a chevron is. focus.js needed no fix and now has the test 98548a35's done condition asked for and never got: swap a region holding a focused input, assert focus and the caret survive. Five cases, all green. htmx is deliberately not loaded -- what focus.js owns is the pair of events, and raising them to test it would be testing htmx.