Port phase 7: cut over and delete the TypeScript math

[wormeyman]

Delete the TypeScript math in one step, now that all three planets render from
WASM.

Two test populations, handled differently

src/noise/ is imported by 155 of the 213 spec files, so this is the
largest single piece of the migration and it must not be a bulk delete.

• The 98 specs that read a fixture. Their assertions were rebuilt in Rust as
  each phase landed, so removing them is a subtraction rather than a gap. Check
  that claim per file rather than assuming it.
• The ~57 specs that import src/noise/ and read no fixture - memo
  behaviour, catalog shape, request routing. Inventoried one at a time. Some
  move to cargo tests, some become parity or seam tests, and some are testing an
  implementation that no longer exists and simply go.

Also here

• Re-decide whether islands/ and preview/tiling.ts cross into Rust, with
  a real end-to-end number for what they then cost as a share. Before the port
  they were under 1%; after a 13x speedup on everything around them, that share
  necessarily grows, and the point of deferring the decision was to make it on a
  measurement rather than a projection.
• Re-measure the CI shard balance. Do not quote one run: three runs of
  identical spec files have spanned 294-469s on this repo, a 59% swing.

Done when src/noise/ holds orchestration and catalogs only,
pnpm run verify is green, and the deployed app renders all three planets from
WASM with pnpm run verify:deploy confirming the build that shipped.

Parent: #215

[wormeyman]

The islands/ and preview/tiling.ts decision, measured

Measured in Chrome against the dev server, not under vitest. That matters
here for the reason #275 records: only one arm of this split pays vitest's
per-module transform (#267 puts it at 3.7x on this very graph), so a vitest
measurement would have compared the harness rather than the code. Two page
loads, min of 7 passes each.

Box (-3000, -3000) to (3000, 3000), seed 123456, 834 candidates from 51,753
samples:

part	run 1	run 2	share
surveyIslands - noise-bound	133.7 ms	136.9 ms	96.0 - 96.3%
chainComponents - pure	5.1 ms	5.0 ms
largestRectangle - pure	0.4 ms	0.3 ms
pure total	5.5 ms	5.3 ms	3.7 - 4.0%

preview/tiling.ts - no, and not close

• planTiles is 0.00 ms per render (256 tiles for 1024x1024 at tileSize
  64). It is O(tiles), called once, and rounds to zero at the resolution the
  app renders.
• stitchTiles has zero src/ callers. Its own comment says so: the app
  blits tiles straight to the canvas so they appear progressively, and this
  exists for consumers needing a flat buffer, "notably the tiled-equals-untiled
  gate". It is test infrastructure that never runs in the app.

Neither is on a hot path. The file stays in TypeScript.

islands/ - not wholesale, but ONE file is a blocker

The framing in the issue is "does islands/ cross into Rust". Measured, that is
the wrong unit, because the directory splits cleanly:

• cellSurvey.ts (171 lines) is the only file with a real value import of
  the noise chain - makeFulgoraStack from ../tiles/fulgoraCatalog - and it
  evaluates stack.cells.cells(x, y) per position. This is the 96%.
• chainGraph, fullChunks, largestRectangle (279 lines) import nothing
  at all. islandMask imports one colour constant. Together they are the 3.7%.

And this is a blocker rather than an optimisation. islands/ never touches
the engine - grep returns nothing - so the shipped Fulgora island finder (#207)
runs entirely on the TypeScript math this issue deletes. Removing
src/noise/expressions/ and src/noise/tiles/ breaks it. Rust already has
Fulgora cells from phase 3 (#223), so the work is giving cellSurvey an
engine-backed cell source, not porting 1,051 lines.

What this changes about the plan

• preview/tiling.ts: closed, stays TypeScript. No measurement needed again.
• islands/: rewire cellSurvey.ts to the engine before any deletion
  touches expressions/ or tiles/. The other ~880 lines stay TypeScript.
• The speed win is a side effect rather than the reason. At the ~2.46x the
  engine measures in-browser, a survey would go from ~137 ms to roughly 55 ms -
  worth having, but the reason to do it is that the finder stops working
  otherwise.

Related: the inventory is also well past this issue's numbers - 243 spec
files rather than 213, 180 importing src/noise/ rather than 155, of which
14 are already WASM-facing, 109 read a fixture, and 57 read none.