T8/P4: AICC facilitator sidecar — constrained LLM over the K1 registry (no-fabrication by construction)

[mzargham]

Spec §10 P4 on branch feat/t8-concept-definition-app. The LLM is a UX affordance over the existing facilitation API (eBike doctrine): it translates natural language into whitelisted tool calls and renders results back; it gains NO affordance the API lacks.

Traceability:

• K1: only registry tools are offered/executable; unknown tool requests refused + audited → tests/unit/test_aicc.py
• K5/LLM.1: cds_queue_add is a MECHANICAL dead-end — after queuing unsecured canon, further write tools that turn are refused → test_unsecured_canon_escalates
• K2: cds_commit is never offered to the model — committing stays human
• ADR-8 BYO-LLM: OpenAI-compatible triplet (CDS_LLM_BASE_URL / CDS_LLM_MODEL / CDS_LLM_API_KEY), operator config; stdlib HTTP client, no new deps; ScriptedBackend seam keeps the suite green without credentials
• T8a: llm-ergonomics eval (skip-if-no-creds): scored no-fabrication + construction-order adherence

Closes only when feat/t8-concept-definition-app merges to main.

[mzargham]

P4 landed in b426c3a: AICC loop with mechanical guards (K1 refusal at the boundary; queue dead-end enforced — test_unsecured_canon_escalates GREEN; cds_commit never offered), ADR-8 BYO-LLM triplet via stdlib HTTP (no new deps), /chat on cds-serve with in-band executed/refused/escalated reporting, and the T8a llm-ergonomics eval (tests/eval, skip-if-no-creds — run with the triplet set). Suite: 318 passed, fully green. LIVE eval + the no-fabrication-baiting LARP pass need a real endpoint triplet — ready to run the moment one is configured.