Two things differ between machines and only one of them is width: a 4K panel
has more pixels, while a 125%/150% display has the same logical pixels holding
LESS, because every label and margin is taller. Breakpoints written as raw
pixels only hold on the machine they were tuned on.
* ui_scale() derives a factor from font height (1.0 at the 15px line the
layouts were measured against) and every narrow-guard threshold is
multiplied by it, so panes fold when the content is cramped rather than
when a number is crossed.
* The window takes a share of the available screen (80% × 85%) with the old
1180×760 as the floor, instead of opening at that size on any monitor.
* Moving the window to another screen re-pins the assistant and re-decides
the fold, since the new screen's work area and scaling may differ.
Found by tools/check_multi_screen.py, which walks 5 window sizes × 3 font
scales:
* At 150%, Schedule was clipped on 1280 and 1366 screens and the window's
own minimum grew to 1459px — wider than a 1280 laptop, so the app could
not fit at all. The cause was not the lanes: the one-line lane-count
summary in the header reported a sizeHint wide enough to set the minimum
width of the entire window. It now yields first (its text stays in the
tooltip); the window minimum drops 1459 → 752 and holds there at every
scale.
Also: these checkers exited 0xC0000409 from a Qt teardown crash AFTER printing
their verdict. check_probes_bite decides whether a probe caught its mutation by
reading exit codes, so a crash would have counted as "caught" — the round could
have passed while proving nothing. They now flush and os._exit with the real
verdict, and round 5 still catches all six mutations.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Five rounds, each looking from an angle the previous one cannot:
1. text — every "Thay đổi" bullet on the page vs a probe (existing)
2. geometry — check_layout_geometry.py: reading order of the rail, section
order down Monitoring, which side each pane is on, and the size
relationships the design names (hero card, 26px dot)
3. inventory— check_controls_alive.py: every control in controls.json still
present in the built app, attributed to its owning CLASS via the
baseline commit (a file holds several classes, so a per-file
check reported eight dialog controls as missing)
4. eyes — rendered screens read against the wireframes
5. adversarial — check_probes_bite.py: break one feature at a time and fail if
the matching check still passes
What rounds 4 and 5 caught, which 1-3 could not:
* Schedule showed six of seven lanes; the seventh needed a horizontal
scroll. The design says "giữ đủ 7 lane, thu hẹp cho vừa một màn". Lane
minimum width 190 → 150, so 7 × 150 + gaps fits a 1280 window. Round 2 now
measures this instead of relying on someone noticing.
* check_design_parity ALWAYS returned 0. It was a report, not a check: every
probe in it was incapable of failing, so a regression would print on screen
and still exit green. It now exits non-zero when anything is CHUA — which
is what let round 5 detect the two mutations it had been sleeping through.
Also fixed in the harness itself: it rewrote line endings while restoring
mutated files (read_text/write_text translate both ways), and it compared the
tree against "clean" rather than against its own starting state.
All ten checkers green by exit code.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
