Add CarMax task verifiers + judge rubrics

[DEM1TASSE]

Adds the per-task verifiers + judge rubrics for the CarMax mirror (the reviewer
deliverable). Submitted directly to main per the maintainer's request.

• sites/carmax/verify/verify_<0..19>.py — one deterministic verifier per task
• sites/carmax/verify/verify_lib.py — shared helpers, same interface/format as the merged
  sites/merriam_webster/verify/verify_lib.py (--run_dir/--initial_db/--after_db/ --container/--no_llm, Judge → JSON {task_id, pass, reason, evidence})
• sites/carmax/tasks.jsonl — adds verifier_path + judge_rubric to each task

Each verifier is deterministic-first and checks three things: navigation (the agent
opened the relevant page — anti knowledge-shortcut), DB after-state for stateful tasks
(reservation / test-drive / saved / order / appraisal created or removed vs the seed), and
the answer against frozen ground truth (LLM-anchored only where phrasing is brittle,
always anchored on the ground truth).

Validated against the official react agent (agent_demo/agent.py): ran the agent per task,
then ran both graders (verifier + LLM judge) on each trajectory and compared. This caught and
fixed several over-strict verifier checks, and surfaced cases where the verifier is more
correct than a trajectory-only LLM judge (an empty final answer, and a reservation on the
wrong model year — both caught via DB after-state). Task #3 (Tesla Model 3 under 50k mi)
correctly fails as unsolvable.

The CarMax site itself is in PR #24 (still under review); these verifiers pair with it.

[DEM1TASSE]

Superseded by #57, which includes the CarMax site (preserving @Violet24K's authorship) together with the verifiers, so it's a complete mergeable contribution. Closing this verifier-only PR.