feat(smith): accept native image input

[edwin-zvs]

Summary

• recognize explicit local PNG, JPEG, GIF, and WebP paths, including Construct file tokens and pasted attachment paths
• snapshot image bytes into persisted Smith turns so retries and resume preserve the original input
• translate images into native OpenAI-compatible, Anthropic, Gemini, Responses API, and Ollama request shapes
• reject image-bearing turns and history clearly for text-only provider wires such as DeepSeek, with /reset recovery guidance
• document the user flow and durable provider-boundary contract in spec 0214

Tests

• cargo test -p construct-adapter-smith deepseek_dialect_is_explicitly_text_only
• cargo test -p construct-adapter-smith image (11 passed)
• cargo test -p construct-adapter-smith (241 passed)
• cargo build
• git diff --check

Visual evidence

Not useful for this change: it affects Smith input persistence and provider request serialization without adding a new visual UI surface.