chore(deps): bump @tangle-network/agent-eval 0.20.12 → ^0.23.0

[drewstone]

Summary

Bumps @tangle-network/agent-eval from 0.20.12 to ^0.23.0.

Note: the request framed this as a six-version leap from 0.17.3, but the
current main pin is already 0.20.12. The actual leap is three minor
releases (0.21, 0.22, 0.23):

• 0.21 — capture-integrity directives for launch-grade benchmark runs.
• 0.22 — EvalCampaign + replay + always-valid + outcome calibration.
• 0.23 — RL primitives bridging eval to policy training.

API drift

None observed. The only consumer in this repo is
bench/research/webvoyager-agent-eval-loop.mjs, which imports six symbols:

Import	0.23 status
aggregateRunScore	exported, signature compatible (src/run-score.ts)
AxGepaSteeringOptimizer	exported (src/steering-optimizer.ts)
JsonlTrialCache	exported (src/jsonl-trial-cache.ts)
PairwiseSteeringOptimizer	exported (src/steering-optimizer.ts)
runPromptEvolution	exported (src/prompt-evolution.ts)
validateRunRecord	exported (src/run-record.ts)

A runtime resolution check confirms all six load as functions/classes; no
source changes required.

The lockfile shrank by ~23 lines because 0.23 has fewer transitive deps.

Test plan

• pnpm install clean (worked under pnpm 10.23 / Node lockfile).
• pnpm lint (tsc --noEmit) — passes with no errors.
• pnpm build (tsc + asset copy) — passes.
• pnpm test — 127 files / 1543 tests passing.
• Runtime import probe of all six consumed symbols against the installed
  0.23.0 package.

E2E / external-bench scripts (WebVoyager loop, design-bench, etc.) were not
exercised here; they require external API keys and live infra. The static
verifications above cover the surface area touched by this bump.