Night of 10-16: no browser on the laptop, GitHub's computers instead

2026-10-16 21:00 to 2026-10-17 00:50 MDT, Claude, while Matthew was away and then asleep. No crash since the 20:53 freeze.

The rule tonight

After the 20:53 freeze (the crash log's 35th row), Matthew's word: no browser on the laptop. Every browser run went to GitHub Actions instead: the Dress Up branches' tests, the pictures, page screenshots. The laptop ran one pinned job at a time (taskset -c 0-3 nice -n 10, dune -j 4): builds, node test runs, ffmpeg.
Learned: GitHub's ubuntu-24.04 image has Chrome but no ffmpeg, and a job-level if cannot read the matrix.

^ticket(181), the dashboard moved to matthewhammer.org/hazel/dashboard/, in the homepage's repository. dashboard/site_publish.py commits it there with git's plumbing (never touching the checkout), at most every ten minutes, as the homepage's Pages build takes about two. The old addresses redirect with their #ticket-N. The homepage's checkout is now usually behind its origin: pull before pushing there.
^ticket(182), a gallery of these notes at matthewhammer.org/hazel/work-notes/: a card per note with a picture drawn for it, a summary, its next steps and the chips and hover cards. A new note needs its entry in work-notes/gallery.json and a picture in dashboard/gallery_art.py, then notes_gallery.py.
^ticket(183), the pictures of 2026/10/17: Dress Up 0.0.94, with a new folder of the friend climbing each place's ladder (18 pictures), made on GitHub's computers. A control run of 0.0.90 found GitHub's Chrome draws what the laptop's drew, though not byte for byte (264 of 349 the same, 73 off by a few anti-aliased pixels, 12 caught at another moment of an animation). The workflow is on kids-choice-games' archive-183 branch; it can be started by hand once it is on main.
^ticket(164) closed on Matthew's word, after his 🧪 runs of 0.0.94 on the Ubuntu laptop (Intel i7) and the iPad.

^ticket(156), the trial merge of #2469 into #2635 (trial/2635-up-to-date, 9e7258b55b): its 11 new CI failures are fixed. Key.mk read an optional tagName; OutlineRename's module path; and two places where #2469's cell code had never met #2635's syntax: a run of hint "…" test … end tests (the mvu docs) and a do x <- e in block (splice-row.hz). Fast checks now fail only on the 10 expected tests (DefStaticsParity per member, Editing.TableDecorations). Full tests were still running at 00:50.
^ticket(18), views in the worker, coded with Matthew's defaults (perf/2635-views-in-worker-costly, ee885708ac, pushed for CI, no PR): only the costly livelit views (8 ms or more on the page) on the screen run in the worker. The tests pass; nothing is measured yet, as that needs a browser.

^ticket(160) settled on GitHub's computers: the on-screen controller costs nothing we can see. Walking the forest three times each way, holding the on-screen arrows averaged 55.7 frames a second, and the game controller with it hidden 57.4, with a spread of 5.7 between its own runs. No shadow, layer or see-through change mattered. An earlier note of mine said it cost about 5 frames a second; that run never showed the controller (a connected game controller hides it) and sat inside the noise. Only the kids' devices can decide now.
Branches built and green on GitHub, each waiting on Matthew's review: collisions-176, flash-177, postcards-43, diagonals-144, speedlog-166, bgtab-167, measure-179, objects-180, and ci (the tests' workflow). The next Dress Up report, ^du_report(3), is drafted on report-003, not pushed.

Review and merge the Dress Up branches above (then ^ticket(178), splitting the game into files, can start), and publish ^du_report(3) (the draft, on report-003).
^ticket(18): a daytime browser run to measure Full Body's main thread and Kids' Choice's release, against #2635.
^ticket(160): the forest on the iPad and the Ubuntu laptop, held arrows against keys or a controller.
^ticket(183)'s workflow onto kids-choice-games' main, if days should be made on GitHub from now on.

^ticket(156): read the trial branch's Full tests, and HeadlessRows there.
^ticket(18): measure the views in the worker in a browser, by day, and choose the threshold.
Merge the reviewed Dress Up branches, one at a time, each with its version.