Module post_turn_verify 

Post-turn verification — “generate first, heal second”.

Alternative to mid-decode Webizen Sentinel gating: run the LLM at full resident speed (like Ollama), then verify and self-heal the completed draft against the quant-graph fact table and a lightweight CML-shaped claim extract before finalising the turn.

Why this exists

Mid-token governance (Phase-8 rings) is architecturally important for hard fail-closed signals, but continuous logit inspection is not what makes Qualia slower than Ollama — the GPU GEMV path is. Post-turn verify still earns its keep: it recovers quality from aggressive INT4 without taxing every token, and produces an auditable HTML/CML surface for the principal.

Pipeline

1. LLM emits plain draft (no mid-decode interrupt).
2. Extract crude claims / capital-style facts from draft + prompt.
3. quant_graph_grounding repair when high-stakes needles mismatch.
4. Emit VerifiedTurn with plain final text + HTML presentation + CML Turtle.

Structs

VerifiedTurn

Result of post-turn verification / self-heal.

VerifyCheck

One atomic check against the local fact / ontology surface.

Functions

maybe_verify_turn

When FastVerify (or quant-graph post path) is active, heal draft → final presentation.

return_html_as_text

Env: return HTML as the primary text field from the agent when FastVerify is on.

verify_and_heal_turn

Full post-turn pass: graph ground + claim checks + HTML/CML packaging.