Правило приёмки скриншотов было про место: снимать только в CI. Обоснование
измерено — съёмка в другом окружении переписывает файлы без содержательных
изменений, в #231 два из девяти на 7–8 байт, набор с беты все девять. Но
держалось правило на комментарии, а не на механизме: кандидат проверялся на
самосогласованность и принимался целиком, ни разу не сравниваясь с тем, что
лежит в репозитории.
Цена видна на #390: правка типов, которая физически не может сдвинуть
пиксель, потребовала прогона workflow, а затем правки одиннадцати полей
манифеста руками.
Теперь правило про доказательство, и оно то же, что у golden с #334: среда
доказана, если каждый кадр, который менять не собирались, совпал с
закоммиченным байт-в-байт. Расхождение растеризации спрятать нельзя — оно
задевает все кадры с текстом сразу. Снимать можно где угодно, включая WSL;
принять получится только оттуда, где кадры воспроизводятся, и перестанет
получаться в тот день, когда обновятся шрифты.
Остальное следует из того же правила: намерение объявляется
--expect-change, необъявленное расхождение останавливает приёмку,
объявленное без расхождения — тоже (ложная декларация обесценивает список),
заменяются ровно объявленные файлы, а тотальная перерисовка требует
--no-witnesses --reason, и причина уезжает в манифест.
Частый случай закрылся сам: ничего не объявлено, все кадры совпали —
принимается один манифест, руками ничего писать не надо.
Проверено шестью сквозными прогонами на подделанном артефакте, не только
юнитами: идентичный кандидат, необъявленное расхождение, объявленное,
молчаливая декларация, тотальная перерисовка без причины и с ней.
Issue: #401
User-Visible: no
CODE-REVIEW-400-r1 Medium: the registered mutant edited a comment, not
the order — it could not reproduce the regression AC1 exists to catch.
That is the same defect class this issue is fixing elsewhere, in my own
guard.
The order is now HANDLE_PAINT_ORDER, a named constant, because it IS the
hit priority rather than an accident of where the blocks sit in the
template. The mutant flips that constant, so it reproduces exactly the
behaviour the audit found.
Also: smoke_furniture picked the SE corner as handles[3], an index that
silently depended on the old paint order — CI shard 3 went red on four
checks. It now selects by role (corner handles, third of four), which is
what the test actually means.
User-Visible: no
Issue: #400
(1) Corner and edge handles carry the same hit radius (1.8 % of the
view), so on furniture narrower than 4·hr — a 40 cm cabinet — the two
circles overlap and whichever is painted last takes the tap. Edges were
painted last. Corners are now, because a side handle scales one axis
while a corner scales both, and the object is small exactly when
proportional resize matters most. The visible beads are unchanged.
The audit called this 'proportional resize becomes unavailable'; the
measurement says otherwise and the spec records the correction: the
corner centre lies outside the edge circle, so the corner was reachable
— its area was halved, not lost. A polish, not a bug, and worth fixing
because it is one line of ordering.
(2) Alignment guides in the devices mode excluded the dragged marker by
_drag, which has been null there since #74 moved device dragging into
_deviceDrag. So the marker being moved was among its own candidates.
Nothing looked wrong because a point always matches itself within
tolerance — the guide was drawn from the marker to itself, visually
identical to an honest one, and the smoke asserted only guides() >= 1.
The smoke now compares the candidate lists with and without the drag and
demands exactly one removed entry.
(3) The 38 settings-help strings stay in the initial chunk, and that is
now a recorded decision rather than an oversight: measured 2 654 B gzip,
0.9 % of the ceiling, against splitting a synchronous dictionary in two,
a second request on first hint, and a second source for the key type
derived from en.json (#391). docs/ARCHITECTURE.md says so, with the
number that would justify revisiting it.
Both mutants run by hand: reverting the paint order reddens the 40 cm
probe while the 160 cm one stays green; restoring _drag reddens the
guides smoke.
User-Visible: yes
Issue: #400
CODE-REVIEW-399-r1 High: the test named after AC5 called
installsPythonDeps on string literals and never executed the directory
walk it was supposed to protect. The reviewer showed what that costs:
restoring the old hardcoded pair of real names and dropping a third
workflow with unpinned installs into .github/workflows left all ten
checks green — the exact scenario AC5 describes went undetected.
The walk is now a function taking the directory, so the test can run it
for real: it builds a temporary directory with three files (a pinned
installer, a workflow that installs nothing, and a rogue one) and
asserts on what the scanner returns. Reverting the walk to a list of two
real names now reddens this test, verified by hand.
The mutant is sharpened accordingly: it substitutes the two-name list
instead of a one-name list. The old form failed on an unrelated
assertion about directory size, so it proved nothing about the scan
itself — while the two-name form is indistinguishable from correct code
on today's tree, which is what makes it the likely regression.
User-Visible: no
Issue: #399
Three claims a green backend used to make, each slightly wider than the
truth — and #392 happened in exactly that gap.
The frontend pin said 20260826.1 next to homeassistant==2026.8.3, whose
package_constraints.txt requires 20260729.7: a combination that exists
in no HA release. It was never derived from anything — someone once
picked it. It is now taken from the constraints, the source is named in
the file, and a test holds both numbers together so raising HA cannot
quietly desync them.
ruff's include declared three trees while CI linted one. Narrowed the
declaration rather than widening CI: the debt in scripts/ and
tests_backend/ (56 findings, mostly E402/I001, plus 7 B023 and 5 B017)
has its own cost and its own decisions, and belongs in its own task, not
in a visibility fix. test/lint-scope.test.mjs now compares the two, so
they can only move together.
The pin check skipped a workflow when it found neither the package name
nor the requirements path — and both vanish together the moment someone
returns to Defaulting to user installation because normal site-packages is not writeable, i.e. the gate switched itself off
under precisely the change it exists to catch. It now walks the whole
.github/workflows directory and decides per file by a positive sign: if
a file installs python packages, it must install them from the pins
file. Verified by dropping a rogue workflow into the directory — it
reddens without touching any list.
Three mutants registered and each run by hand.
User-Visible: no
Issue: #399
The guard introduced by #394 matched the literal
sys.modules['custom_components... and therefore never looked at
pure_imports.py, which writes through a variable — the third instance of
the #389 class walked straight past the check created for it.
The guard now inspects the write itself and decides by the key: a whole
literal or the literal head of an f-string is safe unless it starts with
custom_components (that is how tests register homeassistant.*, hp_pure.*
and houseplan.trails); anything else — a variable, a concatenation,
setdefault/update — counts as a violation whenever the file is able to
name the package at all, i.e. contains a custom_components. literal. A
file that never names the package cannot poison it through a variable,
so restoring a snapshot stays legal.
load_pure now removes what it registered. Removing its own name is not
enough: relative imports pull neighbours in, so junction_limits leaves
wall_segment_model and coordinate_canonicalization behind. It removes
the whole custom_components difference accumulated during exec_module,
in a finally, and a repeated call still works.
pure_imports.py is a named exemption of the static guard precisely
because that guard cannot see the cleanup — so the cleanup is proven by
an executable test instead, and the mutant pure-imports-stops-cleaning
reddens it. Both mutants were run by hand.
User-Visible: no
Issue: #398
CODE-REVIEW-397-r1 Medium: AC3 named the second reader of the same
value — _loadFromServer via _adoptStructuralResponses — and nothing
exercised it. Adding the scenario turned out to be less mechanical than
it looked, and both obstacles are worth recording:
The reconnect path reads BOTH answers, and a differing config clears the
history for its own reason (configChanged). The fake server now echoes
the config the card already holds, so the check answers the layout
question it claims to answer.
The first version of the probe used a round 0.42, which canonicalization
leaves untouched — the check passed with and without the fix, i.e. for
the wrong reason. The probe now starts from a non-canonical position
(0.024999999999999942, which snaps to 0.025), and the scenario runs
immediately after the write, before any reload can align the two sides.
Verified by removing the fix: five checks red, now including
reconnectKeepsHistory and deleteEchoKeepsHistory. Both were green in the
weaker version — which is exactly what the reviewer's Medium was about.
User-Visible: no
Issue: #397
B3: _persistDevicePlacement sent canonicalizePosition(...) to the server
and left the raw value in _layout, then recorded the fingerprint over
that raw snapshot. Canonicalization is not identity — it snaps to the
lattice — so 39 of 115 pixel-derived coordinates differ, and the next
_reloadLayoutOnly or _adoptStructuralResponses saw its own write as a
remote edit: history cleared, _layout replaced. The old _persistLayout
wrote the canonical value back; the per-device path introduced by #74
lost that line.
M1: the smoke that was supposed to prove AC10 assigned
serverLayout = structuredClone(c._layout) right before the reload —
erasing by hand the very divergence it existed to catch, so it could not
fail. The fake WS already stores what went over the wire; the
assignment is gone and the check now reddens on the unfixed code
(verified: three checks red without the fix, including this one).
Also proven, because the fix touches their neighbourhood: the echo of a
DELETE keeps the history (the branch removes a key rather than replacing
a value), and an in-flight write still wins the merge against a server
answer holding the old position.
One existing assertion was loosened deliberately: undo now restores a
position that may differ from the raw one by the lattice snap (<1e-9 of
the plan). That is the point of the fix — local and server agree — so the
equality is stated to that precision, with the snap size pinned
separately so a real drift would still fail.
User-Visible: yes
Issue: #397
Three findings of the v1.70.0-beta.1 audit, all on the transition path
added by #82, all of the same shape — the new path did not inherit a
property the old one had.
B1: persisting the zoom moved into _settleCameraTransition only, and a
cancellation never settles. Touching the plan mid-flight — the literal
scenario of the issue — froze the shown frame and threw it away; before
which kind it is: the user one (_stagePointerDown) persists the frame
that stays on screen, the eleven structural ones keep writing nothing.
The distinction is now also written down in spec #82 §13, which had one
line for both.
B2: the anchor was read from the presented (lagging) frame while the
zoom accumulated from the target, so a six-notch trackpad series walked
the point under the cursor 17 px away — against §10's own promise. Both
now come from the same state. Spec §10 said to use the presented frame
and to keep the anchor within 0.5 px; those two are incompatible, and
the paragraph is corrected rather than left as a trap.
M2: the feather freeze keyed on the two gesture flags, which an
animated transition does not set, so every tween frame rebuilt the blur
region. It keys on 'the camera is still' now.
Guards: unit tests pin the anchor at 1e-9 across 8/16/33 ms series and
prove zoom accumulation is untouched; the smoke checks the shown zoom is
the stored one, that a structural cancellation stores nothing, and that
the anchor holds; three mutants (cancel-loses-zoom, anchor-from-
presented, feather-thaws) were run by hand and each reddens.
User-Visible: yes
Issue: #396
Matysh