Public voice catalog — drives the voice picker on /listen/preferences.

DERIVED FROM app/voices.py, deliberately. This used to be a hard-coded list maintained beside the real registry, and the two drifted until they shared exactly two entries out of ten:

• The catalog advertised kokoro-warm, kokoro-clear, four openai-* and two eleven-* ids that exist in NO registry, so GET /v1/voices/{id}/sample 404'd on eight of the ten voices it had just offered — you could not audition what you were asked to pick.
• It also HID ten real voices, including the named show hosts (Jessica, Chloe, Wit, Brock, Dex, Sage) and both Spanish voices. Listeners could not choose the voices the platform actually ships.
• Its tier field sold plus and pro voices the router will never reach: router.py refuses to let a metered provider LEAD a chain (test_no_chain_leads_with_a_metered_provider), and ELEVENLABS_MONTHLY_FREE_CHARS is 0. Picking "Rachel (Pro)" silently rendered Kokoro instead — the old docstring said so outright.

One source of truth removes all three at once. Every registry voice has a self-hosted path (Kokoro, Piper or StyleTTS2), so every voice offered here is genuinely free to render; self_hosted states that rather than implying a paid ladder that does not exist.