feat: add scoped read-only audio preview and isolated Dify transport

server/.env.followup-audio.example

@@ -2,6 +2,10 @@
 # Only dify may reuse existing [prescription_ai] base/key. openai_audio needs MODEL; asr_then_llm needs both stage model/key/base triples.
 [followup_audio]
 ENABLED = false
+PREVIEW_ONLY = false
+# IDs only. Empty or invalid diagnosis allowlist denies preview; admin list never replaces RBAC.
+TEST_DIAGNOSIS_IDS =
+TEST_ADMIN_IDS =
 AUDIO_VERIFIED = false
 VERIFIED_PROFILES =
 PROFILE = qwen
@@ -66,3 +70,26 @@ OPENAI_EXTRACTION_MAX_TOKENS = 8192
 # The local Qwen run exhausted 8192 tokens with thinking=true; false produced JSON, not a quality pass.
 # QWEN_EXTRACTION_ENABLE_THINKING = false
 # OPENAI_EXTRACTION_ENABLE_THINKING = false
+
+# Synthetic connection readiness only; never sets AUDIO_VERIFIED or allows clinical adoption.
+QWEN_PREVIEW_VERIFIED_FINGERPRINT =
+OPENAI_PREVIEW_VERIFIED_FINGERPRINT =
+
+# Explicit per-stage protocol. No automatic fallback. Dify uses dedicated App keys only.
+QWEN_ASR_PROTOCOL = openai
+QWEN_ASR_BINDING_REVISION =
+QWEN_EXTRACTION_PROTOCOL = openai
+QWEN_EXTRACTION_BINDING_REVISION =
+OPENAI_ASR_PROTOCOL = openai
+OPENAI_ASR_BINDING_REVISION =
+OPENAI_EXTRACTION_PROTOCOL = openai
+OPENAI_EXTRACTION_BINDING_REVISION =
+# For Dify: ASR_PROTOCOL=dify, EXTRACTION_PROTOCOL=dify_chat, RESPONSE_FORMAT=prompt_json.
+# Every Dify stage needs a nonempty BINDING_REVISION. MODEL is expected identity, not sent.
+# Dify App owns generation parameters: local max_tokens/response_format/thinking are NOT forwarded.
+# Leave EXTRACTION_ENABLE_THINKING absent for Dify; set generation limits in the dedicated App.
+# Changing that App requires a new binding revision and fresh preview fingerprint acceptance.
+
+# asr_then_llm only: curl (default) or openssl_stream (explicit verified-TLS child process).
+# No automatic transport fallback; changing this requires new fingerprint readiness.
+HTTP_TRANSPORT = curl