test(skill): add a portable behavioral benchmark

[nstranquist]

Finding

The portable skill passes ndev skills quality --strict with a 10/10 score and no findings. The command ndev skills benchmark wip-commit has no registered runner.

The product already has extensive unit, race, hosted-platform, and real capture evidence. This issue is for a stable skill-level regression baseline, not a replacement for those tests.

Acceptance

• keep a deterministic runner with the portable skill, under skills/wip-commit/benchmark
• make NDev resolve a runner from the current external repository before its internal fallback, without duplicating the skill package
• use disposable Git repositories and invoke shipped entrypoints
• cover shared and linked-worktree setup, disjoint parallel lanes, reviewed split defaults, explicit single-commit opt-out, foreign-index preservation, dry-run evidence, failed verification, and reconciliation
• include a negative control for each safety guarantee
• write a bounded, versioned JSON report with a score, case results, elapsed time, and zero-AI usage fields
• support a reviewed baseline and write a proposal when the score regresses
• run without network access, credentials, private repository data, or a user-specific path
• keep the fast benchmark distinct from the complete unit, race, and hosted Windows suites

Done evidence

• ndev --json skills benchmark wip-commit succeeds from this repository
• a deliberately weakened candidate fails at least one case and cannot replace the baseline
• Linux, macOS, and Windows runs produce the same semantic case results