A picked raster is now classified from its HEADER BYTES ONLY before anything
heavy happens: src/backdrop-probe.ts parses PNG IHDR (+colour type/tRNS for
alpha), JPEG SOF and WebP VP8/VP8L/VP8X at fixed offsets, never using a file
field as an allocation size; hostile or truncated headers collapse to
'unknown', which warns without numbers instead of passing silently. The
thresholds live in that module as the single calibration point
(WARN_DECODED_BYTES 128 MiB ≈ 32 MP, HARD_DIMENSION 16384 — the browser
canvas cap, DOWNSCALE_TARGET_PX 4096), derived from the desktop-Chromium
matrix now committed as demo/benchmark_backdrop_decode.mjs with a
conservative tablet margin documented in the spec.
The shared pick flow (src/backdrop-pick.ts) feeds BOTH lazy runtimes — the
editor space dialog and the onboarding first-space dialog — so the guard
cannot drift between them, and nothing of it enters the eager View graph.
Warn shows the real numbers and three actions; the reduced copy decodes
EXIF-aware, keeps aspect and alpha (PNG stays PNG, opaque becomes JPEG
q0.9) and flows through the ordinary planFile → upload path. Hard has two
phases with one outcome: beyond 16384 px only Cancel; a failed or timed-out
(10 s) reduce closes with a toast, clean staging and NO silent fallback to
the original the user just declined. SVG never reaches the probe. The safe
path swaps the manual byte-loop base64 for FileReader — half the JS-heap
peak on every upload, byte-identical output (parity asserted in the smoke).
Proofs: header-table units incl. a fuzz set of hostile headers and ±1
threshold bounds; smoke_backdrop_guard on the real bundle — zero decode
calls before the choice, byte parity of keep-original, a real 6200 px
reduce to 4096 for both alpha and opaque branches, cancel-only hard
dialog, both phase-2 failures (reject and hang under the test-only timeout
override), re-pick after refusal, SVG bypass; four registry mutants
(probe-always-safe, alpha-dropped, hard-demoted, phase-2 silent fallback).
Spec anchor corrected alongside: the server plan limit is 8 MB
(MAX_PLAN_BYTES), attachments are the 50 MB path — an 8 MB JPEG is easily
80-160 MP decoded, so the client-side guard stays the primary defence.
Issue: #39
User-Visible: yes
Add Deutsch across all card surfaces and backend flows, backed by the language registry introduced in #62. German loads as a fingerprint-checked page-shared locale chunk so EN/RU remain synchronous and the initial View budget stays intact. Root render gates prevent mixed-language flashes, retry once, and fail open to English. Extend parity, runtime, bundle, browser and visual coverage, plus contributor and user documentation.
Issue: #348
User-Visible: yes
The owner's decision (2026-08-28): optimize is one of the two commands a
client can use to write arbitrary geometry, so it validates its candidate
against the stored document exactly as config/set does — inheritance counted
per rule (repairing a legacy plan with violations still passes; #329 AC10
already proves an honest optimization adds none, so the gate is a no-op for
legitimate flows), while a crafted payload is refused with the stable
junction_limit_<rule> code the except list has been ready for since #329.
The call lives inside the existing executor function, and a successful
optimize refreshes rt.junction_baseline with the candidate's counts so the
next config/set inherits from the cache (#330 §4.2 symmetry).
Import and backup restore stay OUTSIDE the gate on purpose — #329 §3
promises a restore is never blocked. The module docstring stops promising
more than the code does, and spec #329 §5 records the perimeter and the
trade-off explicitly: a crafted import can persist violations, but they are
inherited, never legalised as new ones.
HA tests pin AC1 (crafted spike refused, stored config and rev
byte-unchanged), AC2 (echo-optimize of a stored plan that already carries a
violation passes) and AC3 (the follow-up config/set takes its baseline from
the cache — observed through a recording wrapper). The
junction-limit-optimize-unguarded mutant turns AC1 red through the
backend-test-guard convention.
Issue: #333
User-Visible: no
Red dev caught it ninety minutes after the merge: smoke_plan_drawing_repairs
and smoke_resize_pointer_real_plan went red because the new "a 0° wedge is
always a duplicate" rule refused two ordinary edits — creating a room over
an existing partition ring (#308's legal overlay) and resizing a wall until
it lands on a neighbour's. The premise was wrong at the model level: a
shared wall of two adjacent rooms IS two co-located owner atoms on one line,
so every shared-wall node carries a legitimate 0° pair by construction.
Bisection pinned the exact cut: with only the 0° rule reverted, both smokes
are green again; keys, incidence, the iterative walk and fail-closed stay.
Spec revision 4 records the revert and returns "an exact duplicate wall is
invisible to П1" to the status of a KNOWN LIMITATION — an honest detector
needs owner identity, which is a separate decision for the owner to make.
The zero-wedge mutant is removed with its rule; the .5-tick parity unit now
observes quantisation through valence instead of the retired duplicate
visibility; changelogs drop the over-promise.
Issue: #331
User-Visible: yes
Six normative cuts, both mirrors symmetric (spec revision 3):
- §2.1 node keys quantise to 1e-7 with the repository's canonicalisation
formula (sign·floor(|v|·1e7+0.5)/1e7, -0 normalised) — toFixed(6) keys
split one node into two on floating debris and produced two false П4
refusals on a legitimate resize (reproduced: -1e-8 vs 0). Node pairs
within 2e-7 of each other (raw coordinates) are ONE node, and the
node-to-wall incidence uses the same quantum.
- §2.2 a ~0° wedge IS a violation: two rays leaving a node the same way are
a duplicated or overlaid wall (a butt joint yields 180°, never 0°) — the
worst degenerate case was invisible while 0.5° was refused.
- §2.3/§2.4 the wall run is an iterative edge walk over the collinear
component: no recursion (10 000 atoms answered, not RangeError), no
silently dropped fork (the old .find lost every branch but the first),
O(E) by construction, and collinearity is measured against the BASE
segment's axis so an arc of 0.9°-per-atom pieces cannot pose as one wall.
- §2.5 an exception while judging the CANDIDATE refuses the write with the
junction.limit_check_failed toast (fail-closed, as the #278 guard); the
baseline branch stays fail-open by design and the smoke proves the
asymmetry by breaking only the second call of the deterministic pair.
- §2.6 the python mirror narrows its except on the candidate side only:
a genuine migration bug (TypeError) surfaces as an honest WS error, while
a previous-side bug keeps the wide "no baseline" fallback — the two AC6
cases pin the asymmetry so swapped sides turn a unit red.
Parity fixtures gain the new boundary classes (debris node, duplicate wall,
collinear fork); four new mutants pin the filter, the key precision, the
dropped branch and the fail-open hole.
Issue: #331
User-Visible: yes
r2 M-r2-1: AC6 now mirrors AC5's structure with two explicit cases — a
candidate-side TypeError is an honest WS error, a previous-side TypeError
falls back to "no baseline" and the unrelated write passes. An
implementation with swapped or missing asymmetry turns at least one of the
two units red. L2: the risk wording follows §2.3's component-sum phrasing.
Issue: #331
User-Visible: no
r1-H1: node keys quantise with the repository's canonicalisation formula
(sign·floor(|v|·1e7+0.5)/1e7) — native Math.round and Python round() part
ways on .5 ticks, the exact parity lesson coordinate-canonicalization
already encodes. r1-M1: the incidence threshold becomes 2e-7 over raw
coordinates, which the spec's own example (1.02e-7) actually satisfies; the
known valence undercount on neighbouring quanta is stated in §3. r1-M2: the
narrow except applies to the candidate side only — a previous-side migration
bug stays a "no baseline" fallback, symmetric with §2.5. r1-M3: the branch
walk is an O(E) edge traversal of the collinear component, not a
combinatorial DFS; AC3 gains a 100-fork case. r1-M4: the USER-GUIDE limits
section documents the new refusal toast. L1: §1 opens with the user
sentence.
Issue: #331
User-Visible: no
Quantised node keys with -0 normalisation and node incidence at the quantum,
zero-degree wedges become visible, an iterative maximal-branch wall run, arc
collinearity measured against the chain base, fail-closed candidate checks,
and a narrow except in the python mirror. Every reproduction in §1 was
verified by execution on current dev after #330.
Issue: #331
User-Visible: no
H2: the benchmark budgets were calibrated on the author's sandbox with a
1.14x margin — the review runner measured tsFullCandidateMs at 169-171 ms
against a 100 ms ceiling. Budgets now keep the spec's 2-3x allowance over
the SLOWEST observed machine, and the benchmark runs as a step of the
Validate perf job on every push (it needs no browser and no bundle), not
only inside the weekly mutation gate.
M1: the promised AC1 backend test exists now and does what AC1 means: it
patches validate_junction_limits with a thread-recording wrapper inside the
real HA harness — on the event loop that would be MainThread — and proves
the verdicts survived the move (a clean write is accepted, a write adding a
spike is refused with junction_limit_angle). Spec revision 4 rewrites AC1
around this invariant instead of a fragile millisecond assertion.
M2: §4.6 equivalence is now behavioural on both sides (three boundary
fixtures each: as-is counts equal through-migration counts, TS and python),
and the parity suite gained the §7 boundary fixtures (exact 15°, exact
20 cm, the thickness-step filler run, exact 5 cm).
H1 was already closed by 7513f93d (the review ran on the previous HEAD):
check-docs is green on this tree — the screenshots and their manifest come
from one capture run.
Issue: #330
User-Visible: no
Writing the §5 benchmark honestly exposed a cost the point measurements of
П1-П4 could not see: П5 recomputed the full junction topology and masonry
union PER ROOM — 4.2 s per candidate on the benchmark grid. Revision 3 adds
the shared-pass cut (one topology pass per check, the union only when
multi-wall nodes exist, and the resize path reusing its own preflight
artifact) with the measured numbers. Budgets in §5 already assumed the fix;
they are now achievable and proven by the passing benchmark.
Issue: #330
User-Visible: no
r1-H1 was right twice: the rev cache never touched the candidate's migration,
and П4 is architecturally quadratic. Profiled instead of guessing: the money
is not in deepcopy (3 ms) but in _atomize (663k distance calls), and it runs
even for a document that already carries the current catalogue — 815 ms
python / 69 ms TS for a no-op migration. Two new cuts follow: §4.5 bucket
index for П4 (prototype: 372→44 ms, identical verdicts) and §4.6 current-
version documents are used as-is (an explicit revision of the "both sides
through one migration" wording, guarded by a new parity case: v9 input gives
the same verdict with and without migration).
r1-H2: AC4 now rests on the new benchmark that actually exercises the
junction code; benchmark_safe_resize is named as a non-proof. r1-M1: §9
adds the mandatory i18n/touch/risks/release sections.
Budgets in §5 are recomputed from measured post-fix prototypes with a 2-3x
allowance, including an honest row for the one-off cold legacy case.
Issue: #330
User-Visible: no
Four cuts, zero verdict changes: the ws_config_set validator chain moves to
the executor, the previous-document violation counts are cached by
config_rev, П3 builds its node index once per check in both mirrors, and the
frontend baseline is cached per config epoch. A new benchmark with budgets
pins the class of regression (O(n²) returning) in CI.
Measured on dev 2c20f2dc: a 576-atom plan costs 2.8 s in the HA event loop
per config write today; the spec's acceptance bar is ≤50 ms of loop time.
Issue: #330
User-Visible: no
AC10 was asserted, never shown. Optimize runs alignAllToGrid and
repairNearAxisRoomWalls, which move nodes by fractions of a centimetre, and
none of П1-П5 carries a margin wider than the grid step in general — so
"obviously true by construction" was not available.
Two units, both counting violations the way the write barrier does (each side
through commitWallSegmentModel first):
- the owner's fixture in legacy storage — the inherited apex is there before
Optimize, and no rule's count grows after;
- the П4 boundary — two rooms exactly 5 cm apart, where snapping could have
pulled a node under the limit, stay clean.
The first test asserts the baseline actually carries a violation, so it cannot
pass by measuring an empty plan; violationsByRule fails loudly if the space or
its catalogue goes missing, for the same reason. Spec revision 7 records the
proof and the other three review answers.
Issue: #329
User-Visible: no
custom_components/houseplan/junction_limits.py repeats П1-П4 for the write
barrier in websocket_api, counting per rule so an inherited violation still
round-trips, and raises JunctionLimitError with the stable code
junction_limit_<rule>.
П5 is deliberately not mirrored — it judges the rendered wall bodies, and a
second mitre/inset pipeline in Python would drift more dangerously than the
rule it guards. Optimize stays outside the check for the same reason migration
and import do: it repairs existing geometry.
test_parity_with_the_frontend_checks feeds identical fixtures to the TS
functions and to this module and demands the same verdict, so the two
implementations cannot silently diverge.
Issue: #329
User-Visible: no
Measured what Resize itself already forbids: a room cannot be squeezed below
30 cm (two 15 cm walls), so П3 and П5 are unreachable through shrinking and
the gate merely fails closed there. П4 IS reachable on a fine grid, so the
smoke drags a real handle on a 2 cm grid: two rooms 10 cm apart, a 6 cm pull
would leave 4 cm between foreign nodes, the wall stops at 6 cm and exactly one
toast names the 5 cm rule.
Spec revision 6 records both the measurement and the two corrections it forces
on AC7a: a dimmed handle cannot express a per-step limit, and the plan is NOT
byte-unchanged — the allowed part of the gesture is a legitimate edit.
Issue: #329
User-Visible: no
П3 measures the WALL, not the catalogue atom: a short filler segment that
compensates a thickness step (owner's fixture, 5 cm = (30-20)/2) is a legal
continuation of a long same-thickness wall, so the rule walks the maximal
collinear run through the shared nodes before judging the length.
Resize stops at the last allowed position and names the broken rule instead
of the generic "geometry cannot be saved"; the Thickness dialog refuses
through its own toast. Both channels are pinned by demo/smoke_junction_limits
plus three mutants (angle threshold, write barrier, degenerate apex bevel).
Issue: #329
User-Visible: yes
Owner report: small serrations remained on the outer edges between the inner
and the outer vertex. Measured on the fixture ring: two ~4 cm steps plus four
micro-vertices at the tip. Their source was the inset contour's two-point
bevel folding into a bow-tie, and the earlier half-plane clip of that fold,
which left a 0.2 cm sliver the boolean union turned into steps. The inset now
ends in ITS own mitre point at a degenerate apex — mirroring the sharp outer
tip — so there is no fold to clip and no sliver to smear: the room ring is
exactly three vertices, every side longer than the half depth. The clip
helper and its cap plumbing are gone. The user-visible wording of this work
already stands in both changelogs from the #329 entry.
Issue: #329
User-Visible: no
Owner correction (chat, 2026-08-27): no flat chamfer at the tip — a plain
sharp apex. Proven by execution on the issue fixture: the outset contour fell
back to a two-point bevel (the 4·h mitre limit against an 87 cm reach) while
the inset contour folded into a bow-tie, and subtracting that fold carved the
V-notches — together they made the trident. Now a degenerate corner (below 15
degrees, inner faces meeting inside both walls) contributes ONE outset point
at the plan's own vertex — no bevel, no metres-long mitre needle — and its
inset is clipped at the convergence line so no fold is subtracted. Plain and
merely sharp pairs keep the full mitre of #310. The write-side limits of
П1-П5 stop new plans from creating such corners at all.
Issue: #329
User-Visible: yes
Sergey Matyunin