Add Skill vs no-Skill behavioral eval comparison

[GeoGeekLab]

Summary

This PR adds controlled Skill-vs-no-Skill evaluation support and strengthens the behavioral suite around second-order engineering risks.

A/B harness

• add --skill-mode disabled|enabled to native Codex and Claude Code adapters
• add repeated runs with fresh workspaces, run indexes, and agent wall-clock timing
• add a case-by-case comparison report without collapsing engineering quality into one score
• keep qualitative must_do / must_not_do criteria outside automatic scoring

Eval correctness fixes

• fix flaky-test by renaming fixture module token.py to token_value.py, avoiding collision with Python's stdlib token module
• add a regression test that materializes the fixture and imports the production module from the isolated workspace
• replace the verification false-positive substring check with final_not_claim_any
  • negated statements such as “not fully verified” are allowed
  • positive claims such as “fully verified” still fail
  • “not only fully verified …” is treated as a positive claim
• add command_no_changes so generator checks can require a true fixed point instead of silently repairing stale generated output
• compile hidden python -c checks during --validate-only without executing fixture code

Stronger second-order cases

The suite grows from 14 to 17 scenarios.

Existing scenarios now include stronger hidden evidence for:

• security-boundary: symlink escape
• concurrent-worker: failure propagation and pending-work cancellation
• database-migration-rollout: backfill, old-writer/new-reader compatibility, and rollback-safe dual writes
• generated-code-change: generator drift / idempotence
• verification-evidence: execution markers proving local and blocked checks were actually attempted
• scope-expansion: a known unrelated failing regression that must remain out of scope

New scenarios:

• security-toctou: file replacement between validation and open
• authorization-boundary: cross-tenant object access
• api-error-compatibility: exception type, code, and message compatibility

Most of these second-order assertions live in hidden deterministic checks rather than visible fixture tests, so the fixture does not simply reveal the intended answer.

Validation

Latest PR head passes:

• quality: Python 3.10
• quality: Python 3.12
• quality: Python 3.14
• package: success

Repository validation, compilation, unit tests, eval-schema validation, package checks, and release invariants all pass.

The negation-aware claim check is intentionally a small deterministic token-window rule, not an LLM semantic judge.