Part of #1592

Scope

Implement the basic non-streaming chat provider for Google Gemini using the generateContent endpoint.

Files to Create

• crates/zeph-llm/src/gemini.rs -- GeminiProvider struct, request/response types, LlmProvider impl

Files to Modify

• crates/zeph-llm/src/lib.rs -- add pub mod gemini;
• crates/zeph-llm/src/any.rs -- add Gemini(GeminiProvider) variant, delegate_provider, list_models_remote, set_status_tx, with_generation_overrides
• crates/zeph-core/src/config/types.rs -- add Gemini to ProviderKind, GeminiConfig struct, gemini_api_key to ResolvedSecrets
• crates/zeph-core/src/bootstrap/provider.rs -- add ProviderKind::Gemini arm in create_provider, "gemini" arm in create_named_provider
• crates/zeph-core/src/instructions.rs -- add ProviderKind::Gemini to detection_paths match
• src/init.rs -- add Gemini option to provider selection wizard
• src/acp.rs -- add Gemini arm to ACP model switching

Key Implementation Details

• API key via ?key= query param (NOT Authorization header)
• System messages extracted into systemInstruction field (separate from contents)
• Role mapping: User -> "user", Assistant -> "model"
• Model name in URL path: /v1beta/models/{model}:generateContent
• Parse candidates[0].content.parts[0].text from response
• Track usageMetadata for last_usage() (promptTokenCount, candidatesTokenCount)
• Context window lookup by model name (1M for 2.5-pro/flash, 2.0-flash)

Request/Response Types

GeminiRequest { contents, system_instruction, generation_config }
GeminiContent { role, parts }
GeminiPart { text?, inline_data?, function_call?, function_response? }
GeminiGenerationConfig { temperature, top_p, top_k, max_output_tokens }
GeminiResponse { candidates, usage_metadata }
GeminiCandidate { content, finish_reason }
GeminiUsageMetadata { prompt_token_count, candidates_token_count, total_token_count }

Acceptance Criteria

• provider = "gemini" in config.toml works end-to-end
• Non-streaming chat produces correct responses
• API key resolved from vault (ZEPH_GEMINI_API_KEY)
• Error handling: rate limit, invalid key, model not found
• context_window() returns correct value per model
• name() returns "gemini"
• supports_streaming() returns false (Phase 2)
• supports_tool_use() returns false (Phase 3)
• supports_embeddings() returns false (Phase 5)
• supports_vision() returns true
• debug_request_json() mirrors actual API payload
• Init wizard includes Gemini option
• Unit tests with wiremock for success/error paths