[timzifer]

polar was the weakest of the six

An evenly spaced ring of evenly filled wedges reduces, at the resolution a
perceptual hash works at, to "a grey annulus" whatever the cells say. Its
marks scored far closer together than they looked, because the information that
distinguished them lived entirely in the high frequencies a coarse comparison
averages away.

So the radial profile now carries identity too:

• ring widths are drawn, not uniform — the variant stream gives a release
  family its radial rhythm, the patch stream re-proportions it;
• each ring has its own density, so a figure has a brightness profile from
  the middle outwards rather than one uniform grey;
• a centre device — disc, ring, rosette or star — fills the hole, which is
  the most visible position on a radial mark and was previously wasted;
• ring and sector counts span a wider range.

	before	after
weakest relation	14	16
patch minimum	14	16
minor minimum	16	16
major minimum	18	22
global floor	6	8
mean over all pairs	43.1	47.7
intra- vs inter-major mean	41.9 / 43.3	44.4 / 48.3

The family signal is stronger as well: the gap between intra-major and
inter-major means widens from 1.4 bits to 3.9.

Four ideas that were measured and rejected

Recorded on the renderer, because they are all the same reason and the next
person will have the same ideas. On a radial figure, anything large enough for
a coarse comparison to see is large enough to dominate it:

Idea	What the measurement said
Heavy outer rim, archetype stream	Every version of a major shares one silhouette. Cohesion up, patch minimum 21 → 12.
Same rim, patch stream	Drowns out the symmetry that separates majors: major minimum 22 → 14.
Overall size on the variant stream	Separates minors (12 → 15) and pushes patches together by as much (21 → 12).
Independent per-ring rotation	Breaks the fold symmetry a major is recognised by: major minimum 22 → 14.

What remains is the configuration whose weakest relation is strongest — which
is what matters when the guarantee is a floor rather than an average.

A wasip1/wasm leg in CI

The library is pure Go with no syscalls beyond reading a corpus file, so it
ought to run anywhere Go runs. WebAssembly is the cheapest way to prove that
rather than assume it, and it is a genuinely different target: no FMA
contraction, 32-bit pointers, a sandboxed filesystem. If the marks come out the
same there, the suite's float tolerances are doing their job.

wasmtime is pinned and fetched from its release tarball rather than pulled in
as a third-party action. The race detector does not support wasm, so that leg
runs without it; the native legs cover it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KwWthYNYf1egKNFrFS83Dm