Grade expression_in_range against its oracle in Rust (#227)

[wormeyman]

First of the cheap-tier delete-with-loss ports. tiles/expression_in_range.rs
has six tests and all six assert shape on hand-written inputs, so nothing on
this side ever read oracle-expression-in-range.seed123456.json. Deleting
test/expressionInRange.spec.ts with the rest would have left that fixture
committed, version-pinned, and graded by nothing.

The gap is not theoretical. Planting the pre-f32 f64 arithmetic in the builtin
itself - which is the exact defect this fixture was captured to catch - leaves
all six existing tests green, and only the three sweeps added here fail.

Five tests, from the spec being retired:

• the bounded 1-D sweep, expression_in_range(20, 1, (x/1000), -0.5, 0.5)
• the unbounded 1-D sweep, which also pins that an infinite maximum does NOT
  clamp: in-range values run past 1, peaking near 2.5, and a clamp would
  silently kill sand-1's coastal boost
• the 2-D sweep, which pins that dimensions combine by min rather than by
  product or sum
• the f64 arithmetic rejected against the same fixture, so the three above
  cannot be recording whatever the implementation happens to do
• sand-1's asymmetric, unbounded shape against the hand-derived formula

All three sweeps are EXACT - every position of 121, 121 and 162. That matters
because the assertion these replaced started life as a toBeLessThan(8e-3)
ceiling which the wrong implementation passed comfortably; the real residual is
about 9.5e-7, so the ceiling was some 8400x too loose.

Two deliberate departures from the spec:

• The f64-rejection count is frozen at 34 of 121 rather than the spec's "more
  than 10". A number that moves is a finding.
• The sand-1 case keeps a 1e-6 tolerance rather than tightening it. The builtin
  rounds every step to f32 and intermediates like 1 - 1.2 are not
  representable there, so -1.0 comes back as -1.000000238418579. Asserting the
  exact f32 result would mean recomputing the function inside the test and
  checking it against itself. Bit-exactness is the three sweeps.

One thing worth knowing for the deletion: the fixture stores the unbounded
sweep's maximum as the STRING "inf", so it cannot be read as a number.

Claude-Session: https://claude.ai/code/session_01UVcbv1pAhPUoCC6aBwZUtg

Co-authored-by: Claude Opus 5 (1M context) noreply@anthropic.com