`expr::<T>` on a float value whose width differs from T is a width conversion
(`let n = a::<f32>` narrows f64 -> f32; `n::<f64>` widens). Rides the EXISTING
value-turbofish surface — no `as` cast, no invented Convert trait. Zero new
floor leaves; ISA row-set stays 23.

- ck_types: `synth_local_ref` — a float local pinned to a float width synthesizes
  as the TARGET width (the conversion result), so `n == 12.0::<f32>` ties
  same-width. Type-directed on `Primitive::is_float`, never a `match type_name`.
- ck_lower: new scalar-tier `ExprKind::FloatConvert { from, to, value }`; the pass
  wraps the ref (source width = local's checked width, target = the reference's
  checked width); a same-width pin folds to identity. Decomposed into
  `lower/convert_lower.rs`.
- ck_codegen: `apply_float_convert` emits a single Cranelift `fdemote` (narrow) /
  `fpromote` (widen) at the primitive-op tier — no leaf. Float scalar concern
  extracted to `value::emit_float_scalar` to hold the fn cap (no cap raise).

Gate: `tests/corpus/float_convert` narrows 12.0->f32, widens back (bit-exact
f64->f32->f64 round-trip), narrows 0.5->f32, each proven against a pinned
exactly-representable literal -> exit 42 freestanding. Verified to discriminate
(12.0 -> 12.5 yields exit 1).

Honest deferrals (recorded, not faked): f8 has no IEEE-8 machine type; f16's x86
conversion needs soft-float libcalls (__truncsfhf2/__extendhfsf2) a freestanding
binary cannot link — both are an honest backend error. f32/f64 are the native
SSE widths the gate proves. Annotated `let x: T = e` is a separate pre-existing
codegen gap (fails for Int too) — the corpus uses manifest inference.

Gates: fmt clean, clippy 0, cargo test --workspace exit 0, container exits 42
(discriminates), ISA 23, caps unchanged.

@@ -0,0 +204,4 @@
+}
+
+/// Search the test executable's ancestor directories for a sibling file `name`.
+fn find_near_exe(name: &str) -> Option<PathBuf> {

@@ -0,0 +95,4 @@
+    }
+
+    // (5) low KB.
+    let size = std::fs::metadata(&binary).expect("binary metadata").len();