Narrow the incoming coordinate to f32 (#309, #191)

[wormeyman]

Closes #309. Closes two thirds of #191 and updates it with what is left.

The noise machine holds its coordinate values at f32, so an op's x and y are already f32 when it reads them, and the input_scale multiply is an f32 operation on both operands.

Settled by measurement - from a fixture we already had

The plan was a far-field capture at |x| >= 65536, where the 1/256 MapPosition grid stops being a subset of the f32 grid. It would have worked and it was unnecessary.

A derived coordinate leaves the f32 grid right next to the origin. fulgora_basis is a multioctave read at Fulgora's distorted coordinate wx = ox + wobble_x * wobble_mask, computed in f64 and therefore off the f32 grid at 55 of that fixture's 101 positions. Scoring sumOctaves with and without its incoming narrowing:

incoming coordinate	exact	worst residual
narrowed (what ships)	101/101	exactly 0
un-narrowed (planted)	81/101	7.0333e-6

Twenty positions discriminate. That is also the check #191 asked for in its own words - "a caller that passes a derived coordinate" - which Fulgora has satisfied since it landed, unnoticed.

What changed

• basisNoiseExpr narrows x and y before the input_scale multiply, which is what the Rust already did. basisNoiseExpr narrows the coordinate product at a different point than the Rust port #309 was the two ports disagreeing about this. offset_x is 0 at all six call sites, so whether the game narrows before or after that add is unreachable, and no third variant is invented for it.
• variablePersistenceMultioctaveNoise narrows y, in both ports. It narrowed x only as a side effect of the f32(x + offset_x) add; y had no add and so was silently multiplied in f64. Both ports carried the identical asymmetry, which is exactly why tier 2 could not see it - they agreed with each other while both disagreed with the game.
• quickMultioctaveNoise needed nothing. Narrow incoming coordinates to f32 in quickMultioctave, variablePersistence and basisNoise, once a derived-coordinate caller can prove it #191's text is stale: it already narrows both coordinates at quickMultioctaveNoise.ts:192-193. Read the code, not the issue.
• basisNoise itself is deliberately unchanged. Its remaining direct callers (nauvisShared.ts:133, startingPatches.ts:185, regularPatches.ts:164) are unported Nauvis chains where the whole expression is un-narrowed, not just the coordinate. Planting the internal narrowing leaves all 26 of their specs passing, so no committed fixture discriminates it. That belongs to Port phase 6: Nauvis and the five overlays #226, scored layer by layer under the greedy-accept rule.

The pin is inverted, not deleted

the two ports agree off the f32 grid now asserts 0 of 74 diverging where it froze 32, and reverting primitives.ts reproduces [ 'hairlineCracks', ...(31) ] - planted, not assumed.

Its anti-vacuity is not optional: "nothing diverges" is exactly what a sweep evaluating nothing would report, so the test also asserts the off-grid window's folds differ from the on-grid window's on all 74 fields.

Also: the three tier-2 review points from #310

• The field selector moved into fmw-noise. VulcanusParity sits beside VulcanusStack, so elevation_fields and temperature - which no render path reads - are private again instead of pub on a library type solely for a test. Pure code motion, checked as such: tier 2 and tier 3 pass unchanged. It moves engine.wasm by 142 bytes from inlining, which is a reminder that a wasm diff is not by itself evidence of a behaviour change.
• The slider anti-vacuity guard is frozen at its measured 50 of 74 instead of > 40, matching every other count in that file. A floor of 40 would sit through nine fields silently dropping their slider dependency.
• verify:rust's cost is recorded as a range, 1m44s to 2m48s - Build Vulcanus tier 2 and pin the divergence it found (#225) #310 measured 1m44s on CI against the 2m48s in CLAUDE.md, on unchanged Rust. Same spread the test shards show, so widen the figure rather than replace it.

Two lessons recorded in CLAUDE.md

The second is bigger than the bug: fold the fixtures you already have before capturing more, and "no fixture can grade this" is a claim about the fixtures you looked at. It was true of every Vulcanus fixture and false of a Fulgora one.

pnpm run verify green: 2143 passed, 4 skipped, verify:rust included.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AA9K55yddm5D7DPuSJareA