Fix/benchmark

[gitsad]

What does this PR do?

Corrects the OpenUI arm of the generative-UI format benchmark. Their published
benchmarks/system-prompt.txt is byte-identical to their repository but no longer matches the
library they ship: openuiLibrary.prompt() from @openuidev/react-ui@0.13.5 emits 17,431 chars
against the artifact's 13,080, drops ## Examples, adds ## Action - Button Behavior and
## Final Verification, and introduces the Modal and Action components. We were benchmarking
a prompt no integrator receives, so the scored arm now uses the shipped library prompt with the
schema regenerated from the same version.

All 270 generations were re-run rather than adjusted after the fact. Also pins the A2UI Python
setup, which lived in a /tmp venv the OS cleans, leaving the committed benchmark unscoreable on
a clean machine.

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Packages Affected

None. Only benchmarks/genui-formats (private, never published) and pnpm-lock.yaml.

Checklist

• I have read the CONTRIBUTING guide.
• My code follows the existing code style (biome format and lint pass on all 19 files).
• I have added or updated tests that cover my changes — pnpm verify gained probe-adapter
  handling so the superseded arm skips the fixture and cross-validation checks that cannot
  apply to it.
• All tests pass — pnpm verify green; pnpm score reproduces every number offline.
• Type-checking — the benchmark package has no tsconfig.json, so pnpm typecheck does not
  cover it. tsx compiles it on every run and pnpm verify exercises all six adapters.
• Changeset — not needed, no published package affected.
• MDMA schemas — n/a.
• Sensitive fields — n/a.

How to Test

1. cd benchmarks/genui-formats && ./a2ui-python/setup.sh — installs the A2UI SDK from its pinned
   commit. Required before scoring, and the thing this PR makes reproducible.
2. pnpm verify — all six adapters, five gates. Prints All adapter checks passed.
3. pnpm score && pnpm report — re-scores the committed 1,647 generations offline and regenerates
   results/REPORT.md. Free, no API key.
4. Optional, needs OPENROUTER_API_KEY:
   pnpm bench --formats openui-v2 --models google/gemma-4-26b-a4b-it --scenarios contact-form/minimal --k 1 (~$0.01).

Results

"Every time" rate, five repeats per scenario:

Format	Opus 5	GPT-5.6-terra	Gemma-4-26B
MDMA	94.4%	83.3%	94.4%
OpenUI Lang	94.4%	94.4%	50.0%
json-render	83.3%	72.2%	38.9%
A2UI (transport)	83.3%	61.1%	44.4%
AGenUI	38.9%	72.2%	77.8%

Effect of the correction on OpenUI:

Model	Published artifact	Shipped library	Delta
Opus 5	83.3%	94.4%	+11.1pp
GPT-5.6-terra	100.0%	94.4%	-5.6pp
Gemma-4-26B	55.6%	50.0%	-5.6pp

Reviewers should know this correction did not favour us. It gained OpenUI 2 scenarios on the
flagship and cost them 1 on each cheaper model, averaging to no change across the ladder, and
OpenUI now ties MDMA at 94.4% on Opus 5 where we previously showed them at 83.3%. The published
artifact was not systematically unfair to them; it scored better on two rungs of three.

The finding is written up as section 8 of results/REPORT.md, including the prompt diff and the
before/after comparison.