Smoother busy rushes; dashboard lists that fit

[No0ne558]

Smoother busy rushes, and dashboard lists that fit

The main thread was profiled during the stress test's triple-load rush, here and on the store computer (a Raspberry Pi CM5). This fixes what it found.

Speed

• Floor plans: each table's status used to rebuild every open check and the floor plan, for every table, on every change. Now all tables are worked out at once, and the table widgets watch a change counter (checksRevision) instead of converting the whole open-check list for each table.
• Kitchen and expo screens: these looked through every check closed that day, on every change. They now skip closed checks with nothing left to make or run. A recalled ticket, a check closed again after a reopen, a menu or settings change and End of Day make them look again.
• Bumping a ticket: saving a learned prep time saved all the settings and told every screen the settings had changed. Prep times are now saved quietly, at most once a minute.
• Handheld updates:
  • written to JSON directly instead of through Qt's JSON classes; a new test checks the output against Qt's for every kind of value;
  • tip shares worked out once per round instead of once per screen;
  • the version history built once instead of on every update;
  • a round that took long waits twice as long before the next, so updates never take more than about a third of the store computer's time.

Results

• Here, at 5× load: worst pause 2,080 ms → 193 ms, none over 250 ms; handhelds' slowest 1 in 20, 363 ms → 98 ms.
• On the Pi, 13-hour triple-load day: worst pause 4.2 s → 1.7 s; pauses over a second 136 → 31. Every check balanced, and End of Day took 0.3 s.

The Pi results above come from the old stress setup, which ran 13 people's screens on the store computer itself. A Pi run with the realistic setup below hasn't been done yet.

Tests

• Stress test: now realistic. Only the manager works at the store computer; everyone else uses a screen or handheld of their own, over the network.
• Demo test: the card-name check waits for at least ten card payments. It failed by chance just after midnight UTC, when all of the day's few card payments were keyed in.

Fix

• Dashboard: the lists (on the clock, best sellers, running low) now scroll inside their cards. On a busy day, names were running off the bottom of the screen.

556 tests pass, and the same suite passes on Qt 6.8, the Pi's version, apart from the version-history check, which can't run without the git history.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Weg3awu6SARK7jYrtvDfuw