fix(gateway): Gaia's 16 Oct issues, DAG barge-in and streaming TTS, realtime close, build and test fixes

[dittops]

Summary

Root-cause fixes for every Gateway issue in Gaia's 16 Oct report, plus the follow-ups found while verifying them. Each item was checked against the same inputs as Gaia's original release build (mock vendors, a real waav-infer, an auth responder and a real LiveKit 1.13.10 server).

Reported issues

Issue	Root cause	Fix
DAG TTS never delivers; failures hidden	select! dropped the completion receiver; errors only logged	Poll the receiver in place; typed DAG error messages
LiveKit room calls 404	livekit-api encodes the path's slashes; LiveKit 1.13 rejects it	Vendored livekit-api with a literal path
${VAR} sent literally	Conversation path passed the raw key	Named credentials, resolved per LLM leg
Inline DAG can send the operator key anywhere	Env expansion for client graphs; SSRF bypasses	Credentials bound to hosts; SSRF and DNS guard
Debug logs print API keys	Derived Debug and tracing spans record keys	Redacting Debug; spans skip secrets
API secret disables the auth hook	Secret mismatch returned 401 before the responder ran	AUTH_MODE chain: api_secret or responder
Responder 200 {} admits callers	Empty body treated as an identity	Identity required, else 502
STT 5 s and LLM 60 s limits, no knobs	Constants	Env and config knobs (the DAG LLM knob is now honoured too)
/metrics is public	No policy on the route	MetricsPolicy: token, separate listener or 403
Parent-directory .env is read	dotenv walks parent directories	./.env or --env-file only
Segmented STT off unless allowlisted	Default allowlist ignored explicit requests	An explicit request counts as opt-in
Denoiser never runs on /ws	/ws never called it; LiveKit frames too short	Per-session streaming denoiser, opt-in; warms up without blocking ready

Follow-ups

• DAG barge-in and LLM→TTS pipelining. Each turn runs as its own task on the streaming executor; barge-in cancels it and clears TTS, and a newer turn supersedes an older one. The TTS node opens one provider per turn and speaks each sentence as the LLM writes it. Live: a barge-in cut the old answer within ~1 s and the new one was answered in full; first audio ~1.6 s earlier.
• DNS rebinding on webhook / HTTP DAG nodes: they now use the connect-time guarded resolver.
• Realtime close on a slow client. Teardown dropped the socket with the client's unread pongs buffered, so the kernel sent a TCP reset and the client lost its remaining audio and the 1011 close. Teardown now keeps reading while the backlog and close go out; the grace for the close reply runs in the background, so the admission slot is released at once.
• Transcripts in logs. Content fields are <redacted> in the stdout log unless WAAV_LOG_CONTENT=1 (OTLP export unchanged).
• Failed setup closes the session. A refused credential or failed DAG / conversation / LiveKit setup ends the session with a 1011 close instead of continuing as STT-only.
• DAG definition lint reports unknown keys, including inside arrays (nodes[1].voice_ID).
• OpenAI TTS URL no longer doubles /v1.
• aarch64 build. Vendored webrtc-sys builds the NVIDIA codec on x86 only, so a fresh aarch64 build with CUDA installed succeeds without CUDA_HOME/clang workarounds; Cargo.lock regenerated.
• Tests run on a clean checkout. The realtime and voice-span test files generate their credential and JWT keys at run time instead of reading bud-auth's git-ignored *.pem fixtures. Two breaker tests now open their 30 s breaker after the gateway is up, and the slow-client test sends more than the kernel can buffer.

Behaviour changes to note

• Transcripts are no longer in the stdout log by default (WAAV_LOG_CONTENT=1 restores them).
• A session whose LLM credential is refused, or whose DAG / conversation setup fails, now closes with 1011.

Test plan

• 644 / 644 targeted tests (DAG, security, realtime relay and translate, voice spans, LiveKit, credentials, denoiser), with ORT_DYLIB_PATH set (turn-detector tests need ONNX Runtime)
• DAG live verification against mock vendors and a real waav-infer (256,800 bytes of Kokoro audio delivered)
• Barge-in and pipelining live check
• LiveKit setup failure closes 1011 in 0.06 s (original stayed open 20 s)
• Fresh aarch64 build with real CUDA installed

🤖 Generated with Claude Code