CLI structural integration + CVA6 full struct_enum benchmark verification

[metasmile]

Goal

Integrate StructuralEnum into the ev verify CLI path, replacing
expand_all + evaluate_all with the structural enumeration pipeline.
This makes the 73x-1000x speedup available to all CLI users.

Also: verify the CVA6 full (33M) struct_enum estimate by running the
actual benchmark.

Background

The library already has StructuralEnum (structurally filtered iteration)
and build_runtime_checks (runtime-only constraint evaluation). But the
CLI ev verify still uses the old pipeline:

Current: expand_all → evaluate_all → reporter    (~32s for CVA6 full)
Target:  StructuralEnum → runtime_checks → reporter  (~32ms for CVA6 full)

Tasks

1. Add cva6_xif_ref fixture bench to benches/bench.rs
2. Run cargo bench -- struct_enum/cva6_full to get actual measurement
3. Update cva6.qmd, ibex.qmd, index.qmd with confirmed numbers
4. (Optional) Add --structural flag to ev verify CLI

Expected result

Metric	Current estimate	Verified
CVA6 full struct_enum	~32 ms	TBD
CVA6 full speedup	~1000x	TBD

[metasmile]

Branch 42-cli-structural-integration created from main for the CLI structural integration work.

[metasmile]

CLI structural integration delivered on branch 42-cli-structural-integration (commit da2ecb9).

Changes:

• src/verify/compose.rs: raw_total_combinations computes the cartesian product size from field domains without enumeration, sharing the overflow guard with expand_all.
• src/verify/evaluate.rs: evaluate_structural returns the raw total plus the evaluations of the structurally valid subset (StructuralEnum + evaluate_all).
• ReporterCapable::report gains a total parameter; reporters derive failed = total - passed, keeping counts authoritative when the evaluation list holds only the valid subset. The simulate channel is unchanged (passes evaluations.len()).
• ev verify routes through evaluate_structural by default.

Measured (release, this machine): CVA6 full fixture 10.4 s -> 0.21 s, with the same counts (33,554,432 / 196,608 / 33,357,824) on every committed fixture.

Tests: the two speed-ignored CLI tests now run and assert summary counts; two lib-level tests in tests/structural_enum.rs pin evaluate_structural against expand_all + evaluate_all. Suite: 96 tests, none ignored. run.sh asserts the CVA6 full and R4 counts in CI (the 33M skip is removed).

Note: --json on the 33M fixture now emits the 196,608 valid rows only, which is why the un-ignored tests assert text summary counts instead.