0001 — Provider registry with fallback chains and strict mode

• Status: Accepted (backfilled 2026-09)
• Drivers: every AI layer (LLM / images / TTS / avatar) has cloud, stock-photo, and self-hosted implementations of very different reliability

Context

A render touches four independent providers. Any of them can be unconfigured (no key yet), unreachable (laptop asleep, GPU box busy), or failing at runtime. Teachers need "some video, always"; lecturers with a cloned voice or a paid image budget need "exactly the configured provider, or nothing".

Decision

providers/factory.py is the single registry. Each layer resolves as: configured provider → configured fallback (if usable) → a hard floor (gTTS / text cards / no avatar). Strict mode (settings.strict, --strict) converts every fallback into a StrictModeError abort, and runtime failures honour it per-slide too (providers/base.py). Unavailability at startup (missing key/base_url) and failure mid-run are reported distinctly, and slide-stream doctor explains the why (see doctor.py).

Consequences

• Adding a provider is one class + one registry line + one list_* blurb.
• Silent quality degradation is possible (e.g. elevenlabs → gTTS mixed voices in one video); the run summary warns about fallback counts and nudges toward doctor (cli.py _doctor_nudge).
• Strict mode must stay opt-in per run/config; the free floor is what makes the hosted demo and first-run experience work with zero keys.