token-plan profile advertises speech.synthesize but every voice fails with [cosyvoice:] Engine error [411]; no discoverable voice IDs for the default TTS model

[ParallelEntrepreneur]

Summary

On the built-in token-plan profile, bl speech synthesize fails for every model/voice/format
combination, while image.generate and video.generate succeed on the same API key. The CLI-side
problem is that nothing in bl makes this discoverable:

1. bl config show lists speech.synthesize in api_key_capabilities for this profile, so the CLI
   presents TTS as supported.
2. --list-voices for the profile's own default TTS model returns no list, so there is no way to
   learn a valid voice ID from the CLI.
3. --list-voices --model cosyvoice-v3-flash does print a catalogue, but every ID from it fails,
   and the model itself 404s on this endpoint. The catalogue is therefore misleading on this profile.

Environment

bl --version	1.25.0
skills version	1.25.0 (bailian-protocol frontmatter)
npm latest	1.25.0
profile	token-plan
base_url	https://token-plan.ap-southeast-1.maas.aliyuncs.com
default_speech_model	qwen-audio-3.0-tts-plus
OS	macOS 27.0, arm64

Reproduction

bl speech synthesize --text "Every product needs a backend." --voice Ethan --language en \
  --format mp3 --out out.mp3

Actual

[Model: qwen-audio-3.0-tts-plus] [Voice: Ethan]
Error:       [cosyvoice:]Engine error [411]: TTS speak operation failed
Status:      HTTP 400 (InvalidParameter)
Exit code:   1

Expected: either an MP3, or an error that names which voice IDs are valid for this model/profile.

What was tried

• Voices Ethan, Serena, Cherry, Chelsie on the default model — all Engine error [411]
• CosyVoice IDs longtian_v3, longxiaochun_v3, longcheng_v3, longanyang on the default model —
  all Engine error [411]
• --format mp3 and --format wav — identical
• With and without --instruction, --rate, --language
• --model cosyvoice-v3-flash → Model not exist (HTTP 404) on this endpoint, although
  --list-voices for that model prints 26+ voice IDs

Image and video generation on the same key and profile succeed, so this is not auth or quota.

Request IDs

• f1292503-a91f-47b0-841c-aef030aff2e8
• c6094826-a2a2-4d8a-9d78-23559139e184
• 69407cc8-34b6-46ce-942c-25f143782b86

Suggested fix

Whichever way the service behaves, the CLI could avoid the dead end:

• Do not advertise speech.synthesize in api_key_capabilities for a profile where no voice works, or
• Make --list-voices return the voices valid for the active profile and its default model, and
  scope the cosyvoice-v3-flash catalogue to endpoints where that model exists, and
• Surface Engine error [411] with a hint that names the likely cause, as other commands do.