Simplify ROB bench with immutable scenario data

[zhoubot]

The ROB system bench repeated large conditional trees for fifteen stimulus and expected-result fields. Replace them with one immutable typed Table of 284 scenarios, using existing Struct defaults and a full-width phase clamp. The source shrinks from 6,228 lines / 258,859 bytes to 371 lines / 39,366 bytes.

Preserve all 284 transactions, including the 256-event deterministic random stream; the final row keeps flush=1 through the entire remaining u64 domain. Phase initialization/wrap, eight ordered DUT inputs, seven guarded checks, seven unconditional logs and registration order remain unchanged. No new API, compiler change, DUT/oracle rewrite, cycle reduction or timeout increase. Refresh the example guide and verified module receipt.

Validation:

• Independent proof compares all 4,260 field values and 284 full-domain intervals, including sustained flush, clamp and wrap. Twenty-one malicious mutations are rejected.
• Standalone tests pass 2/2. The module retains its 569-sample oracle and failure/discard/retry probes. The system completes 568 sampling epochs and 3,976 observations across native workers 1/2 and Verilator.
• Complete original and candidate observations agree. Cross-source comparison excludes only changed AST registration/site metadata; event values, specs, epochs, instances and order remain checked.
• Source-import and transformed artifacts reproduce both published source units byte-for-byte. Measured and public runs share identical final IR and complete generated outputs.
• Catalog tests 19/19, changed-file pre-commit and strict documentation passed.

Same-machine LLVM 22 / -O0 measurements reduce bench compilation 6.62→1.45 s, linking 6.48→2.67 s and C++ emission 6.14→2.68 s. Three alternating paired warm native runs have medians 1.111→0.641 s (about 42% lower); RTL runtime is approximately unchanged at 0.036 s. Final IR falls 28.20→11.07 MB, generated C++ 5.51→3.69 MB and RTL 1.17→0.65 MB. README records the complete table and measurement limits.

Full catalog/nightly and platform matrices were not run. The known-phase proof does not claim injected X/Z phase equivalence or simulated 64-bit rollover. Remaining migration/scaling work stays in #272 and #265. Intermediate experiments, proof tools and reviews remain ignored under docs/gates/logs/rob-bench-20261010/; runtime tests do not depend on those files.

[zhoubot]

Final candidate bfcd468acb71993da343ed8d05c421964762cdbc: independent code review APPROVE, architecture CLEAR, both required GitHub checks passed. Implementation, independent preservation and review used separate instances.

Evidence preserves all 284 scenarios / 4,260 field values over the complete known-u64 domain, including sustained tail flush and wrap; 21 mutations are rejected. Standalone CTest 2/2, catalog tests 19/19, complete observation comparisons, compiler-stage and generated-output bindings, pre-commit and strict docs passed. Final README reports same-profile phase costs and three alternating paired warm runtime samples, with native median about 42% lower and RTL approximately unchanged.

No compiler, API, DUT, driver, configuration or timeout changes. This is bounded ROB bench acceptance, not full nightly, platform or overall migration completion. Candidate-bound reviews and raw results remain in ignored local evidence.