Give the Vulcanus cliff-blocking tile set one definition per side (#364)

[wormeyman]

Closes #364.

The set of Vulcanus tiles that refuse a cliff - lava and lava-hot - was
written out four times, and nothing checked the four agreed. Two of the four sat
on adjacent lines in fixtures.rs and said the same thing in different
vocabularies, one by enum and one by string literal, so they could drift apart
without either file changing.

Now there is one definition per side, and a test that the two sides agree.

What moved

before	after
cliffs/vulcanus_fields.rs inlined Lava | LavaHot	calls tile.is_cliff_blocking()
fixtures.rs is_lava, the same match again	VulcanusTile::is_cliff_blocking
fixtures.rs want_lava, by name: name == "lava" || ...	resolves through the new VulcanusTile::from_name, then asks the same predicate
cliffCatalog.ts	unchanged, and still the only TypeScript definition

from_name is the inverse of the existing name(), resolved through
TILE_ORDER so the two cannot disagree. It refuses a name Vulcanus does not
place, which is what turns a typo in a fixture into a failure rather than a
silent false.

The two sides are now compared, not just tidied

fmw-wasm exports vulcanus_cliff_blocking_names_fnv1a64(), an FNV-1a 64 over
the blocking tiles' names, sorted and joined by newlines.
test/wasmVulcanusParity.spec.ts hashes VULCANUS_CLIFF_BLOCKING_TILES through
the module's own fnv1a64 and compares.

Three choices there are deliberate:

• Names, not a count. A count cannot tell lava from volcanic-folds, and
  a swap is exactly the drift worth catching.
• The module's own fnv1a64, not a second one in TypeScript. A
  reimplementation is one more thing that can disagree, and its disagreement
  would look identical to real drift.
• Sorted. Catalog order is ground truth for the argmax tie-break, but a
  Set of two strings does not carry it, so hashing in catalog order would make
  this depend on a fact it is not trying to check.

It can fail, and that was watched rather than assumed

Adding VolcanicFolds to the Rust set and rebuilding the module turns exactly
the new parity test red and leaves the other 11 in that file green. A second
test in the spec is the standing control: it hashes the set with a tile added,
with one removed, and with one swapped, and asserts all three differ from the
module's answer. Without it, a bug that made both sides hash nothing would leave
the comparison green while comparing nothing.

The module grew 1,289 bytes, measured rather than assumed

320,092 to 321,381. Stubbing the new export to return 0 unconditionally and
rebuilding gives 320,135, so:

part	bytes
the export symbol and its body	43
the name table, the sort and the hash	1,246
total	1,289

The cost is name()'s 19-arm match pulling every tile name string into the
module. Nothing unaccounted for got linked in. The 512 KB tripwire in
wasmEngine.spec.ts is not close and did not fire.

One thing found on the way, which is NOT folded in

Sweeping the repo for Lava | LavaHot rather than trusting the issue's table
turned up two sites the issue did not list: rocks/vulcanus_placement.rs and
resources/vulcanus_geyser.rs. Both refuse the same two tiles.

They are deliberately left alone, because the three sets are equal by
coincidence rather than by a shared rule. The geyser's own doc comment already
says so - "the forbidden set coincides with the rock overlay's while being
reached by a completely different route" - and the three routes are:

gate	what actually decides it
cliffs	the cliff's collision mask holds water_tile, and tile_collision_masks.lava() sets it
rocks	the rock prototypes' vulcanus_tiles_cold / vulcanus_tiles_hot autoplace lists, whose union is every tile but these
geysers	type = "resource" defaults to the mask {resource = true}, which tile_collision_masks.lava() also lists

Change the tile data on any one of those axes and the three come apart. A shared
predicate would then be wrong in two places at once, and silently. That warning
is now written on is_cliff_blocking itself, where the next person to try the
merge will read it.

Gate

pnpm run verify green locally. The four new Rust unit tests cover from_name
in both directions, freeze the set at the lava pair, and check the exported hash
is computed the long way rather than compared against a magic digest.