bench: dedicated does-depth-help study (deep/residual vs plain)

[davorrunje]

Context

The alternate base-result bake-off (PR #111) is deliberately shallow-only (≤4 effective layers). Depth was deprioritized because a controlled size-ladder study (docs/benchmarks/loan-size-ladder.md) found deep/residual stacks within noise of shallow ones, with the edge not growing with scale. The spec left a dedicated "does depth help" study as a follow-up.

Goal

Run a focused depth comparison — deep/residual vs plain, tuned independently per configuration — to confirm or refute depth-neutrality beyond the single size-ladder dataset, and record the verdict.

Where

• benchmarks/_common/search.py — the flavor set includes *-residual/*-deep (_ALL_FLAVORS); run_dataset / benchmarks/search.py drive it.
• benchmarks/_common/size_ladder_report.py — the existing Δ(deep − shallow) analysis to reuse/extend.
• Results/docs under benchmarks/results/ and docs/benchmarks/.

Acceptance criteria

• Deep/residual vs plain compared with per-config tuning on ≥2 datasets (ideally across sizes).
• A bootstrap verdict (Δ with CI) per dataset, rendered as committed results + a docs page.
• Clear statement of whether depth helps, matches, or is neutral.

References

• PR feat(bench): tuned shallow base result — 4-flavor bake-off + convex_fraction ablation #111; spec docs/superpowers/specs/2026-07-14-alternate-base-result-design.md; docs/benchmarks/loan-size-ladder.md.