fix(tts): read tts.openai.api_key/base_url from config.yaml before env vars

[zccyman]

Summary

_resolve_openai_audio_client_config() was env-only — it read VOICE_TOOLS_OPENAI_KEY / OPENAI_API_KEY and always returned DEFAULT_OPENAI_BASE_URL, ignoring tts.openai.api_key and tts.openai.base_url from config.yaml.

Closes #26175

Problem

Users pointing TTS at a self-hosted or local OpenAI-compatible endpoint (llama.cpp, qwen3-tts, Ollama, etc.) either need a placeholder env var (and still cannot redirect the host) or get no audio at all — even with a fully-populated tts.openai config block.

Solution

Mirror the STT resolver order used in transcription_tools.py::_resolve_openai_audio_client_config:

1. Config — read tts.openai.api_key + tts.openai.base_url from config.yaml via existing _load_tts_config()
2. Env var — existing resolve_openai_audio_api_key() fallback (now also respects cfg_base_url instead of hardcoding DEFAULT_OPENAI_BASE_URL)
3. Managed gateway — existing fallback with prefers_gateway("tts") check

Preserves the existing tts.use_gateway preference — when set, the managed gateway is preferred even if credentials exist.

Updated the error message to mention config as well: "Neither tts.openai.api_key in config nor VOICE_TOOLS_OPENAI_KEY/OPENAI_API_KEY is set"

Files Changed

File	Change
tools/tts_tool.py	Add config-first resolution to _resolve_openai_audio_client_config()

Test Results

tests/tools/test_tts_speed.py ........... 19 passed
tests/tools/test_transcription_tools.py ...... 95 passed
tests/tools -k "tts" .................. 182 passed

Design Decisions

• Uses the existing _load_tts_config() helper (was already defined but unused by this function)
• Config base_url is also respected from the env-var fallback path, unlike the previous code which hardcoded DEFAULT_OPENAI_BASE_URL
• prefers_gateway("tts") check is preserved on both config and env-var paths

[zccyman]

Thanks @NishantEC for the review! Appreciate the approval on the TTS config fix. 🙏

[teknium1]

Thanks for targeting the config-only OpenAI TTS path. The premise is confirmed on current main: tools/tts_tool.py:2533-2535 still only resolves env credentials, while tools/tts_tool.py:1027 calls that resolver before applying tts.openai.base_url at :1032-1034.

Problems

• Current main changed this resolver to return (api_key, base_url, is_managed) in commit b53ba0e18; its caller unpacks three values at tools/tts_tool.py:1027. Salvage the config and env direct paths with is_managed=False.
• Config-only credentials also need to reach check_tts_requirements(): its OpenAI branch calls _has_openai_audio_backend() (tools/tts_tool.py:2494-2496), which currently only checks env credentials or the managed gateway (:2556-2558). The registered tool uses that check at :2868.
• Please add regression coverage for config-only resolution and availability; this PR currently changes no tests.

Suggested changes

• Reconcile the resolver with the current three-value contract and extend the availability predicate to recognize tts.openai.api_key, preserving tts.use_gateway behavior.

Automated hermes-sweeper review.