refactor(evidence): remove dead scoring and canonicalize paths

[drewstone]

The unused RunCritic inferred test quality from command text and treated URLs as task drift.
Its adapter was already removed, and no package entry or production module imports it.
Delete the critic, its three tests, and its stale comment reference.
Keep the public scoring types and the steering optimizer.

Validation exposed an artifact-root mismatch through symbolic links.
Use one canonical root for artifact reads and relative manifest paths.
Add a regression that compares direct and linked roots.
Make the artifact-swap fixture use canonical paths, and let the output fixture drain stderr before exiting.

The worktree adapter also returned paths that its own finalization check rejected through symbolic-link parents.
Return the canonical path of each created checkout so finalization and cleanup use Git's directory identity.
Preserve rejection of substituted worktree locators.
Exercise creation, finalization, verification, and cleanup through linked repository and worktree roots.

Validation on fa167c37817786d3ddfa643804529ce4b026e5c4:

• Full TypeScript suite: 5,746 passed and 3 skipped, both on Linux CI and locally with Node 22 and two workers.
• Python client: 207 passed and 4 skipped.
• Published GEPA compatibility: 45 passed and 1 skipped.
• Official TypeScript optimizer integrations: 2 passed, plus 1 published-GEPA bridge test.
• DSPy compatibility: 18 passed.
• Lint, all three typechecks, build, package validation, and Python distribution checks passed.

CI run