The band test still visited every entity pair. For boundaries of 24 or
more entities, EntitySlideEvents now builds bucket indexes over each
entity's perpendicular extent per enumeration (both sides for vertex
rays, the line-only side for arc-to-line, the stationary curves for
tangency) and visits only the bucket members, ascending, before applying
the same exact band tests. Range queries merge the ascending bucket
lists without repeats. Each extent is the band test's own acceptance
bound widened by a reach of 2*sqrt(QuadraticSlack)*(1 + largest
along-axis separation) for the quadratic tests plus a rounding term;
bucket numbers are monotone in the coordinate, so a query returns a
superset of the accepted entities in their original order and the
emitted events are unchanged. Geometry with any coordinate that is
nonfinite or beyond the band's coordinate limit is not indexed.
SlideBandTests adds large random boundaries (24-160 entities) to the
differential comparison against the frozen pre-filter enumeration,
reruns the band-edge sweeps with decoy entities so they go through the
index (including a far circle reached through the moving offset), and
checks index queries directly. Mutations killed: zero reach, line extent
without span slack, repeats kept by the merge or the sorted path, last
bucket skipped, points just above the minimum treated as empty, curves
in the line index, arc and tangency queries ignoring the radius, moving
index or measurement without the moving offset. Not pinned by a test:
the reach's rounding term (headroom).
Measured on converted Nestbolt jobs (Default engine, hermes, quiet
machine, one run each), before -> after: m-box-tool 47.1 -> 19.2 s,
m-mechanical-mounting-kit 45.9 -> 27.3 s, m-mechanical-cog-washers
74.7 -> 44.2 s, m-mechanical-small-brackets 23.0 -> 16.6 s. The nine
completed jobs keep identical placements.
An axis-aligned slide can only meet an entity whose extent across the push
direction holds the ray. EntitySlideEvents.Enumerate now tests that band
before each vertex ray, arc-to-line pair and curve-tangency pair, and skips
only pairs whose kernels cannot emit. Margins cover RayLine's segment
tolerance (twice Eps of the span) and the rounding of SolveRayCircle's
discriminant. A pair with any coordinate or radius that is NaN, infinite or
beyond 1e150 is always kept, because the kernels' products can then turn
NaN and still emit. Surviving pairs run in their original order, so the
emitted events, and every slide distance, are unchanged.
SlideBandTests compares the full event sequence and FirstBlocking result
bit for bit against a frozen copy of the pre-change enumeration: random
line/arc/circle boundaries at four coordinate scales with offsets landing
on or beside the tolerance band, an arc swept across a nearly parallel
line's band edge, and nonfinite or overflowing coordinates along the push.
Mutations killed: zero span slack, zero quadratic slack, circle-line test
ignoring the radius, with zero margin or a shrunk radius, tangency using
the difference of radii, no coordinate guard, no vertex guard, moving
offset ignored, axes swapped. Not pinned by a test: the span slack's
factor of two and the quadratic slack's absolute floor (rounding headroom).
Measured on converted Nestbolt benchmark jobs (Default engine, hermes,
one job at a time): m-box-tool 70.7 -> 40.7 s, m-mechanical-cog-washers
115.1 -> 70.0 s, m-mechanical-mounting-kit 72.7 -> 44.9 s. The nine jobs
that completed (five M jobs and four B jobs; m-sign-1 still exceeds
300 s) produced identical placements before and after.
aj