test(bench): stage 2 dominance probe — new@low vs old@high

[thewrz]

This was written agentically; verify its assertions:

Slice 8 of epic #152. Design: docs/superpowers/specs/2026-08-13-token-benchmark-design.md, section "Staged plan" stage 2.

Why

The highest-value comparison is new@low vs old@high: effort tiers move tokens by multiples while the refactor moves them by percents, so a dominance result survives a noise floor that a percentage delta cannot (repeat-run token spend varies up to ~30×; arXiv:2604.22750).

What

• 5 paired trials (10 runs), new@low vs old@high, A,B,B,A alternation so provider drift cancels; each round includes the frozen 06d18cf drift-control trial per the pre-registration.
• Adjudicate strictly by the pre-registered dominance rule (median acceptance ≥ AND median blended USD <, non-overlapping IQRs); publish the reference-hit-rate analysis (the ≥90%/0% siting rule) regardless of the dominance outcome.
• Go/no-go record for stage 3.

Acceptance

• 10 trial records + drift controls in the ledger; void-trial rule applied as pre-registered
• Dominance verdict stated against the pre-registered rule, not a post-hoc one
• Reference-siting findings reported (misfiled / dead-weight references named)
• Go/no-go decision note committed

🤖 Co-authored by Claude Fable 5.