Semantha prototype: blind evaluation, ablations, and go/no-go decision

[cuzzo]

Run raw/normalized/facts, 256/768, clustering/similarity, path/name leakage, shuffled null, lexical/identifier/structural baselines; capture blind labels and compute every design gate. Acceptance: one bounded tuning round and explicit invest/search-only/rerun/abandon decision.

[cuzzo]

The baseline/ablation metric machinery, seeded evaluators, deterministic null shuffle, decision gates, and review fixture schemas are implemented. Keeping this open because the real EmbeddingGemma ablations and blind human dispositions cannot be run until the gated model is authenticated; hashing baseline output is explicitly not semantic evidence.