Add verified evaluation harness for /query/verified

[IgnazioDS]

Motivation

• Provide an automated evaluation harness for the verification endpoint /query/verified to validate claim-level groundedness and evidence integrity.
• Produce deterministic verification checks when AI_PROVIDER=fake so eval results are stable for CI and local smoke tests.
• Emit both JSON and Markdown reports for auditability and debugging under scripts/eval/out/.
• Keep the existing /query eval behavior unchanged and reuse the existing debug/source helpers.

Description

• Added a new dataset scripts/eval/golden_verified.json seeded with answerable and insufficient-evidence cases for verification testing.
• Implemented the runner scripts/eval/run_eval_verified.py which posts to /query/verified, validates answer/citation expectations, claim shapes, verdicts, score bounds, evidence integrity, and computes verification metrics.
• Wrote outputs to scripts/eval/out/eval_verified_results.json and scripts/eval/out/eval_verified_report.md, and updated the Makefile with the eval-verified target and README.md docs to include make eval-verified.
• All checks are deterministic under AI_PROVIDER=fake, the runner exits non-zero if any case fails, and existing scripts/eval/run_eval.py runtime logic was not changed.

Testing

• No automated tests were executed as part of this change.
• The change is additive (new files and a new Make target) and does not modify existing eval runtime behavior or existing unit tests.
• Recommend running the stack and harness with AI_PROVIDER=fake DEBUG=true docker compose up --build followed by make eval and make eval-verified to validate behavior locally.
• Created outputs are written to scripts/eval/out/ and the runner is designed to fail the process with non-zero exit if any case fails.

---

Codex Task