test: [P2][resolver] Hand-labeled edge corpus + measured precision/recall baseline

[codeakki]

Problem

We claim ~25–35% in-repo edge resolution but have never measured precision/recall on a labeled set. We can't improve (or defend) what we haven't measured, and any resolver refactor needs a regression baseline.

Proposal

• Hand-label call edges on 3 repos: GraphPilot itself, one mid-size OSS app, one monorepo package.
• Record precision and recall of the current name-based resolver against the labels.
• Commit corpus + scoring script under bench/resolver/ so it doubles as a regression suite.

Acceptance

• bench/resolver/ corpus + script committed
• Measured baseline precision/recall published in bench/resolver/README.md (replaces the estimated 25–35% figure in docs/limitations.md)