Refactor select pipeline scenarios into compact immutable data

[zhoubot]

Replace the select-pipeline bench's expanded conditional expressions with 1,829 immutable typed scenario rows and one clamped lookup. Omitted fields use existing implicit-zero initialization. The source shrinks from 22,903 lines / 1.142 MB to 1,905 lines / 0.293 MB without changing the DUT, native/RTL oracles, four-state tests, configuration or registered 1,829 cycles. No API, compiler or budget changes are needed.

Preserve the original 11-bit phase and wrap, constant tail, five guarded assertions and five unconditional logs. Independent comparison proves all 24,576 field values across all 2,048 phases. Native workers 1/2 and Verilator retain all 18,290 observations. Extended resetless runs preserve the original failure at epoch 4,102 after an identical 20,510-observation prefix; no reset or assertion workaround is introduced.

The cost tradeoff is explicit in the README. On the same Clang22 -O0 profile, bench compilation drops 83.51→21.74 seconds and peak RSS 515.8→362.4 MiB; linked IR shrinks 15.9%. The C++ header grows 11.5%, native build time grows 16.0% and peak RSS grows 19.5%. Two warm runs give native runtime median 42.69→44.24 seconds (+3.6%); RTL runtime remains about 0.33 seconds. This is a source/front-end improvement, not a general runtime speedup.

Validation: existing standalone module/system/four-state CTests 3/3; related units 108/108 and post-catalog checks 19/19; changed-file pre-commit, strict MkDocs and diff checks pass. Source-import and transformed stages reproduce published units byte-for-byte; measured and verified system backends use identical final IR. Add the verified generated receipt/excerpts and link them in the existing catalog/navigation.

Evidence stays local and ignored under docs/gates/logs/select-bench-20261009/. Build-phase measurements are single samples, runtime medians use two warm samples, and results are platform/profile-specific. Full-catalog/nightly and platform acceptance remain open in #272; scaling work remains in #265.