Palette: fold the two rarest interaction types, and hold the UMAP palette to the same floor (#532, #601, #602)

[realmarcin]

Closes #532. Unblocks #199. Implements the decision: fold the two rarest interaction types into "Other".

The constraint this removes

tests/test_network_palette.py stated it in its own docstring:

no nine-colour set survives all three deficiency types at once

The palette shipped with four documented tritan collisions as the price, the worst being CROSS_FEEDING vs MUTUALISM at ΔE 1.8 — the two commonest interaction types in the KB (252 and 229 occurrences), indistinguishable to a tritanope.

Seven colours do survive.

Folding is the enabler, not the fix — worth being precise

The two rarest types are STRAIN_COMPETITION (1 occurrence) and PREDATION (11, against 67 for the next-rarest coloured type). The failing pair was the two commonest. So folding does not directly repair anything.

What it buys is the magenta region those two occupied — and that is exactly where MUTUALISM had to move, because under tritanopia every blue converges on the teal of CROSS_FEEDING. The fold and the re-step are one change.

MUTUALISM #56bbe6 → #9f2ebb.

Measured, with the module's own Viénot–Brettel–Mollon projection

check	before	after	floor
normal / protan / deutan minimum	15.4 / 17.0 / 15.7	15.4 / 17.0 / 15.7	15.0
CROSS_FEEDING↔MUTUALISM under tritan	1.8	68.4	—
tritan collisions between interaction types	2	0	—
documented tritan collisions total	4	1	was ≤4
contrast light / dark	—	5.86:1 / 3.13:1	1.9

The single remaining collision is SYNTROPHY against the grey taxon node — which shape already separates, since taxa are circles and every interaction is a non-circular symbol (#270). The test now asserts zero collisions between interaction types and ratchets the documented total from 4 to 1.

The enum is untouched

All nine values remain — records still use them, and folded types render with the grey "Other" swatch the template already had for unmapped keys. FOLDED is explicit and reasoned, so a newly added enum value still fails rather than silently rendering grey, which is the defect this module was written to catch.

test_folded_types_are_still_the_rare_ones re-derives the counts from the KB, so the justification cannot rot: if a folded type becomes common, or a coloured type becomes rarer than a folded one, it goes red.

What I could not verify, and why the issue's proposal wasn't taken as offered

The dataviz rubric's OKLab validator is not present in this environment (scripts/ in the bundled skill is empty). That is the check that reported the original ΔE 10.3 normal-vision failure, and it should be re-run when available.

The two metrics genuinely disagree on that pair: CIE76 puts #56bbe6↔#57c7ab at 41.1 while the rubric's OKLab put it at 10.3. Separately, the issue's proposed re-step #edab12 → #b8860b fails this module's deutan floor at ΔE 1.5 against COMMENSALISM — so it could not be adopted, and I searched for a replacement that satisfies the gate the repo actually runs.

Mutation checks

reverted	result
MUTUALISM back to #56bbe6	1 failed (tritan collisions)
PREDATION removed from FOLDED	1 failed (enum coverage)
FOLD_HEADROOM 2 → 200	1 failed (usage justification)
nothing	10 passed

Gates

lint 0, validate-all 0, validate-strict 0, 2465 passed. Regenerated HTML for all 309 affected pages is included.

[realmarcin]

Added to this PR: #601 and #602

Working #199 (which #532 unblocks) turned up two more instances of the same pattern.

#601 — the UMAP palette was asserted CVD-safe, not computed

#199's checklist records the UMAP page's colours as "CVD-safe 8 + grey Other". Measured with this repo's own colour code:

	before	after
normal	22.0 (0 pairs below 15)	31.1 (0)
protan	6.9 (4)	16.0 (0)
deutan	10.4 (3)	16.7 (0)
tritan	5.8 (3)	11.0 (4)

The worst pair was LIGNOCELLULOSE/METHANOGENESIS at ΔE 6.9 — the classic red/green convergence under protanopia — and BIOTECHNOLOGY collided with the collapsed grey at 10.4 under deuteranopia, which is a category becoming indistinguishable from the "everything else" bucket.

Re-stepped keeping every hue's semantic assignment (blue rhizosphere, green lignocellulose, red methanogenesis) and separating on lightness, which is what red/green convergence actually requires.

On the tritan trade: one more pair sits under 15, but the worst pair nearly doubles, 5.8 → 11.0. No two categories are now indistinguishable, which matters more than the count. I picked this over a variant scoring 17.4 on the gated metrics precisely because that one dropped tritan's worst pair to 2.2.

Both palettes are now gated by one module — the defect was the same as #588: a rule applied to one artefact and not its sibling.

#602 — the docs gate could not see the UMAP page

check-docs-current ran only just gen-html, but docs/community_umap.html comes from just gen-umap. So that page could drift from its template indefinitely and the gate stayed green, because nothing regenerated it to produce a diff.

Demonstrated by this very change: after editing the template and running gen-html, the gate reported ✅ while the published page still served 'RHIZOSPHERE': '#2a78d6'.

That is the failure mode the recipe's own #442 comment describes — "green in exactly the case where the regeneration it gates on had not happened". The orphan-page check added then fixed one instance, not the pattern. gen-umap now runs there too, and the invariant is written down: every generator whose output is committed under docs/ must run in that gate.

Safe to add — checked that the embedding is seeded, so two runs are byte-identical and this cannot flake.

One parser bug caught in my own work

The first version of _umap_palette matched every hex in the template and dragged in stateColors/originColors — separate schemes for other colour-by modes, never on screen together. It invented three failures. Now scoped to categoryColors, with a shape guard so a parser that returns {} after a refactor fails rather than passing vacuously.