test(bench): stage 3 frontier — effort curves and epic close-out

[thewrz]

This was written agentically; verify its assertions:

Slice 9 of epic #152. Design: docs/superpowers/specs/2026-08-13-token-benchmark-design.md, sections "Staged plan" stage 3, "Sampling, not a grid".

Why

Only if stage 2 justifies it: fill the remaining cells (2 arms × 3 effort tiers) to produce the deliverable — acceptance score vs blended USD, two curves. The matrix is sampled, never exhaustive: default rounds are the drift control plus current main per tier; a new model enters only after the frozen arm establishes that model's own baseline.

What

• Remaining cells at the stage-2-set budget, with drift controls per round and A,B,B,A alternation.
• bench/plot.py: the two curves as a pure function of bench/results/ (no hand-carried numbers).
• Epic close-out: verdicts for Q1 (already met) and Q2 written against the pre-registered rules, with the longitudinal machinery (ledger, fixture versioning, drift arm, per-merge Tier 0) left running as standing infrastructure.

Acceptance

• All plotted points come from the ledger; cross-fixture_version points never share a line
• Curves + verdict committed; epic epic: token-consumption benchmark for the compression refactor #152 closeable with every staging checkbox satisfied
• Sampling policy for future models documented

🤖 Co-authored by Claude Fable 5.