Build the generation-uplift benchmark (test pass-rate with vs without the Brain)

[bejranonda]

The retrieval benchmark harness now exists (PR #125), but the claim that actually backs the positioning — the Brain improves AI coding — is still only a design in docs/VALIDATION.md §2, not code.

Goal: run the same agent on a held-out set of coding tasks that ship with executable tests, twice (Brain-injected vs not), holding model/prompt/seed constant, and report test pass-rate with n + paired difference + CI. No LLM-as-judge.

Acceptance

• Task suite authored independently of the corpus (small internal set, or SWE-bench-lite).
• Agent harness that runs both arms deterministically.
• Metric pre-registered; raw outputs committed.
• Result written into docs/VALIDATION.md (or explicitly reported as null/negative if that's the truth).

See docs/VALIDATION.md "Proposed methodology §2".