ADR 0003: AI provider foundation audit

Status: Accepted for Phase 0

Context

Issue #74 establishes the minimum AI foundation before vocabulary acquisition features are added. This audit records the actual development surface so a later phase does not mistake a planned adapter or prompt for a shipped one.

Findings

• AIProvider currently exposes only suggest_mnemonic(word, context).
• OllamaProvider is the sole concrete adapter. It implements the mnemonic operation, bounded output, timeouts, and prompt/data separation.
• No OpenAI, DeepSeek, or Vertex adapter exists. No configuration value may claim to enable one.
• There is no provider orchestrator, enrichment operation, or vocabulary extraction operation yet.
• The mnemonic prompt protects supplied vocabulary data with delimiters, but there are no enrichment or extraction prompts to evaluate for a target language requirement.

Decision

Phase 0 adds only the provider operations needed to implement an honest, AI-backed extraction path and its explicit test/demo fallback. It must not advertise unimplemented providers or silently substitute a local heuristic when AI is disabled.

Each future provider adapter must implement every operation added to the port or be excluded by configuration. DeepSeek and Vertex remain documented gaps; they are not stubs because an importable but non-functional provider would make the configured-provider contract misleading.

Every enrichment and extraction prompt introduced after this audit must state that examples are written in the requested target language, while retaining the same untrusted-data boundary used by the mnemonic prompt.