test(roundtrip): pin reparse-equivalence and diagnostic text for the fidelity regressions

[e54-bot]

Summary

The five roundtrip-fidelity defects were fixed in the audit batch; this change closes the remaining gap against the issue's acceptance criteria by pinning the outcome contract in the regression tests rather than only the individual assertions:

• Items 1, 4, 5 (rule-name backslash, forward variable reference, opaque settings escapes): now end with assert_emit_round_trip, so the emitted text must reparse to an equivalent program and re-emit to a fixed point — the issue's "reparses to an equal program" criterion.
• Items 2, 3 (fractional/overflow declaration index, duplicate settings block): the tests now assert the diagnostics name the offending construct (integer index, duplicate 'settings' block) instead of only asserting is_err.
• Unverified item (declaration names references can't express, e.g. a.b, a % b): recorded as reproduced and rejected — the diagnostic names the unreferenceable name.

Verification

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test -p workshop-rs --lib — all green
• Re-verified on the CLI: Wait-family emit outputs remain clean; 1.9, 4294967296, -1 indexes and a.b/a % b names all fail with malformed: diagnostics; duplicate settings fails with malformed: duplicate 'settings' block
• git diff --check

Fixes #428