fix(benchmarks): select cached plugin versions numerically - #970

Summary

Resolve cached plugin releases by numeric version components so the agentic benchmark loads the latest installed release. Environment overrides still take precedence; cache selection and activation keep their existing contract.

Reproduction

With both ~/.claude/plugins/cache/ponytail/ponytail/4.9.0 and 4.10.0 installed, the existing lexical ordering chooses 4.9.0. I reproduced this through the public benchmarks/agentic/run.py runner with one cache task and an offline CLI fixture that records the actual --plugin-dir argument. The fixed runner chooses 4.10.0.

The regression check also covers patch versions 4.10.2 / 4.10.12 and preserves an explicit plugin directory override.

Scope and prior work

The numeric ordering idea also appeared in #903, which its author closed for a task-scope misunderstanding. This proposal isolates that still-reproducible bug, leaves checkout activation/provenance/rescore metadata out of scope, and does not overlap the active selftest environment restoration in #717 or unused ARM cleanup in #668.

Numeric chunks cover the current release directories; this deliberately does not implement full prerelease SemVer ordering. The code states when a version parser would be needed.

Verification

• Regression check fails against the unchanged main implementation and passes after the fix.
• Actual public runner red/green replay records the selected cache directory. No real model or provider was invoked; outbound network was denied.
• npm test: 122 passed, zero failures/skips (96 root, 23 Pi, 3 MCP).
• node scripts/check-rule-copies.js: passed.
• node scripts/check-versions.js: passed; all eight version files remain 4.10.0.
• git diff --check: passed.

AI-assisted contribution, independently reviewed and replayed before publication.