The continuity gate cannot see the defect from the owner's 66.json: masonry is
continuous there and the record agrees with what is painted. The record itself
is wrong — a partial resize left 43 steps of a former shared boundary as an
exterior wall while it kept the 20 cm of that boundary, next to 30 cm exterior
neighbours. Thickness followed the key, not the role of the edge.
A width check would not have caught it, and I built one before throwing it away.
It compares the painted body against the record, and here the two agree. On real
plans it also fires where masonry is legitimately wider — columns, junction
influence, abutting parallel walls: 76 to 82 steps measured against 4 expected,
every case legal. A gate that needs explaining half the time is noise.
The defect is expressible in a single state instead: one record whose span is
partly shared and partly exterior. Nobody sets that on purpose. Roles come from
the polygons — a stretch is shared when another room's edge covers it — so the
check needs neither a build nor product code.
Two traps found by measurement, both of which produced false positives. Count
distinct rooms rather than edges: in a corner one room owns two edges, and
counting edges called every exterior corner a shared boundary — 12 and 14 false
positives. And do not sample the endpoints: an endpoint is a node, where a wall
legitimately touches two rooms, and including them reported 95 per cent exterior
on every wall abutting a shared one.
Mutation coverage, stated honestly: the endpoint mutant is killed by the
real-plan test. The rooms-versus-edges mutant survived — once endpoints are
excluded, counting edges gives the same answer on real plans, so that choice is
not load-bearing. I removed the mutant rather than ship a surviving one, and said
so in the code.
Issue: #287
User-Visible: no
Stage 0 of ADR #282. A lattice node is k/240, which has no exact binary
representation, and a stored coordinate is a float. Nobody could say how much of
a real plan is affected, and Optimize promises to remove coordinate noise
without a definition of noise that can be checked.
latticeProfile splits every coordinate of the model into three populations,
because they are three different problems: exactly on a node, near a node but
not exact, and legitimately off grid. The middle one is the defect class behind
\#258, \#279 and the non-converging Optimize; the last one is authored geometry
the current model allows and must not be called a violation.
Measured on the owner's installation: space 1 has 208 coordinates, 33.65 per
cent exactly on a node and 65.38 per cent in the noise class; space 2 has 21.23
against 78.77. The worst deviation is 8e-8 of a step — invisible, and enough to
put a wall key in the neighbouring bucket.
The counterpart is what makes it worth having: every shipped fixture has zero
noise, all of its off-grid values being authored. Our own test data therefore
cannot reproduce this class by construction, which is why the owner finds these
defects and the gates do not. A test pins that property so it cannot drift.
No violations are produced, no gate turns red, and nothing is repaired: what to
do with a vertex 8e-8 from a node is the owner's decision, and this measures its
price first.
Issue: #283
User-Visible: no
The first cut of checkWallKeys compared the stored key against endpoints
snapped to the lattice and called any mismatch a violation. Both halves were
wrong, and measurement says so: wallIntervals reports the query key for the
disputed edge as 0.887500,0.195833@1.5706, i.e. the form built from the
coordinates as stored, and the two owner configurations that differ in exactly
these keys produce byte-identical wall bodies and multi-wall node maps. The
check would have reddened a plan that renders correctly.
Graded now: a drift inside the tolerant fallback's half-pitch reach is an
observation, a key beyond it or one that does not parse as coordinates is a
violation. The threshold is expressed in grid steps with a 1e-3 slack — with a
relative 1e-6 the four identically drifted records of one plan split between
the two classes on their last bits, so the check repeated the very rounding tie
it exists to expose.
visual-matrix leaves KEY_CONTRACT_DEBT: its keys drift inside the reach and now
read as observations. large-house stays — its labels do not parse, and
wallIntervals shows all 80 solid edges resolving to zero thickness (#260).
Issue: #259
User-Visible: no
#258 lost two thickness records to a rounding tie: wallKey quantises the
midpoint with Math.round, and a wall of odd step length has its midpoint
exactly on the tie, where the exact node 83/240 and the stored 0.345833333
fall on opposite sides. The #254 invariants pass on that file — their edge
tolerance is 0.004 and the drift is 0.00417, so the check sits on its own
boundary.
checkWallKeys compares strings, and against endpoints snapped to the lattice:
keying from the raw endpoints flags the healthy state and passes the broken
one, which is what the first formulation in #258 got wrong. Measured on the
owner's before/after pair: 0 findings before, exactly the two artefact walls
after.
Two shipped fixtures write keys off the contract (#260); recorded as a number,
so the debt can neither grow nor be silently fixed.
Issue: #259
User-Visible: no
Claude