fix(mir): evaluate machine when-guards and trap-close the fallthrough

[slepp]

Summary

Machine when guards were type-checked but never evaluated at runtime: synthesize_machine_step_fn's event cascade branched on event-tag equality only, so a guarded transition fired regardless of whether its guard held. I added a guard-check block between the event-eq branch and the arm body, reusing the same self/event binding-env handling emit_machine_step_transition_return already uses, evaluated before any hook or transition-out drop — a false guard now falls through to the next arm exactly like an event mismatch would.

This also closes a latent undefined-behaviour hole: once a guard can evaluate false at runtime, an all-guards-false step with no default arm is a genuinely reachable, user-visible condition rather than the dead code HIR exhaustiveness used to guarantee. I route any machine with at least one guarded transition to a real TrapKind::ExhaustivenessFallthrough trap instead of a bare LLVM unreachable. Guard-free machines are unaffected — they keep the existing (sound) unreachable, since an event match still fires unconditionally there and HIR exhaustiveness proves the fallthrough unreachable.

Verification

• Red-before/green-after: a Closed{code:1} machine with ==42/!=42 guarded arms fired the wrong (Open) arm before the fix; after the fix it correctly falls through to the !=42 arm.
• A true guard still fires (Closed{code:42} → Open), proving the guard is evaluated both ways, not just gated shut.
• An all-guards-false step with no default now traps cleanly (hew: trap in main context: Trap, non-zero exit, no signal) instead of hitting UB.
• New hew-codegen-rs exec tests: IR-shape assertions split into has_default / has_guard / neither, asserting the guard-bearing fallthrough emits the real trap call (not a bare unreachable); execution tests for all four new fixtures including the fallthrough-trap oracle.
• cargo test -p hew-mir and cargo test -p hew-codegen-rs green; cargo fmt and cargo clippy -- -D warnings clean on both crates.
• Full test corpus and o2-differential green in an integrated build alongside unrelated in-flight work; no regressions.

Out of scope

None.