[Air000000]

Summary

• Add an explicitly experimental Hybrid /rag/ask path. Dense remains default; Ticket Agent and the original Dense 0.9 distance policy are unchanged.
• Require both server switches (RAG_HYBRID_SEARCH_ENABLED, new default-false RAG_HYBRID_ASK_ENABLED) plus request retrieval_mode=hybrid and refusal_policy=v5_3_experimental. Reject max_distance on Hybrid rather than silently applying or ignoring a distance threshold. Hybrid defaults to Top14; Dense remains default3/max10.
• Reuse existing Hybrid retrieval and frozen v5.3 pure prompt/strict original-line coverage validators without modifying experiment sources or reading evaluation data. Dedicated Phase B credentials are mandatory for runtime classification, without credential fallback.
• Insufficient/conflicting/conditional/ambiguous evidence returns a fixed refusal without generation. Admitted requests generate from the complete ordered scoped context; require canonical source references drawn from gate-supported IDs. This checks provenance/format, not semantic truth.
• Recheck SQL visibility/content before classification/generation and before publishing. Treat provider, schema, truncation, empty-output, citation and visibility failures as sanitized 503 + failed audit, never normal refusal or silent Dense fallback.
• Preserve nullable Dense distances and separate rerank scores; return/audit experimental policy and validated gate details on successful/normal-refusal outcomes. Add refusal-state filters and complete status-count metrics. No SQL migration.

Verification

• 360 offline tests passed: focused CI suite + auth. New cases cover gating/config, exact frozen prompt and line presentation, tenant/category scope, null and high Dense distances without threshold reuse, unsupported/ambiguous gate outcomes, generation refusal, malformed/truncated/provider-error responses, citation ID restrictions, concurrent document changes and audit/status metrics.
• Real OpenAI-compatible SDK with mocked transports; isolated databases and external network guard. No real model calls, no new fee authorization consumed, no DEV access or public deployment.
• ruff check ., in-memory Python compilation and git diff --check passed. Primary checkout's existing changes left untouched.

Explicit boundaries

This is runtime wiring for an off-by-default experiment, NOT v5.3 promotion or independent semantic-quality acceptance. Historical NO_PROMOTION / FAIL conclusions are unchanged. Valid evidence locators/source IDs do not prove entailment, faithful generation or prompt-injection resistance. It does not replace default refusal or create tickets.

Actual provider execution of the newly added classifier/generator path and controlled answer/refusal acceptance remain pending separate data-egress/cost authorization. The previous 8 embedding + 3 rerank smoke authorization is not reused.