feat(xla): run dense matmul numeric probe

[inureyes]

Summary

• emit a semantically narrow f32 dense-matmul probe with explicit input, accumulator, result, and rounding boundaries
• bind the VMFB cache to the v3 operator numeric contract, compiler identity, MLIR digest, and linked IREE runtime archive SHA-256
• load resident weights and execute the emitted graph through the native auxiliary IREE ABI
• compare identical operands against a row-major sequential Rust decomposition with exact-bit reporting
• add a root example so real-runtime validation receives the established static IREE link recipe

Qualification semantics

This probe is intentionally reported as canonical-decomposition, not production-mlx. The selected IREE CPU contraction kernel remains explicitly unobserved, so a passing report is always production_qualified=false. No model family or capability predicate is activated.

Validation

• cargo-fmt fmt --all -- --check
• cargo check -p mlxcel-xla
• cargo check -p mlxcel-xla --features micro-oracle
• cargo test -p mlxcel-xla --features micro-oracle numeric_ -- --nocapture (21 passed)
• cargo-clippy clippy -p mlxcel-xla --features micro-oracle --tests --no-deps (existing warnings only)
• cargo run --features xla-micro-oracle --example xla_numeric_probe -- local-task repeated successfully through real compile/load/invoke

Observed local-task evidence:

• comparison: exact bits, 4 elements, 0 failures, first divergence null
• operator contract SHA-256: ea04d277cbca9b0d32679ea8dd6cb55b05bbd5503ed1dc99be7a19425c20eea9
• IREE runtime archive SHA-256: 3ab0d811d9ce1f359e7e1f37ecfad7285499cfed5e3c6335265767cfbc7cdf44
• artifact fingerprint: b4461b40278903ca
• transfers: 24 B resident weight, 24 B invocation input, 16 B output
• reported candidate invocation: approximately 2 ms

Compile-time and peak-memory fields are not yet part of the v1 report schema and remain follow-up instrumentation; this slice does not claim to complete #932.

Refs #932