feat(deeptide-rs): multi-protocol backends, themed TUI, reasoning display, local engines

[feizhenhua]

Broadens the Rust deeptide CLI toward zero-cli parity and beyond — across protocols, TUI rendering, headless mode, KV-cache, and local-model UX. Default behaviour is unchanged (the dark theme reproduces the prior hard-coded colours, so every existing snapshot stays green).

Protocols & models

• OpenAI (/v1/chat/completions) + Gemini (generateContent/SSE) backends behind the protocol-agnostic AgentBackend trait, alongside Anthropic — the agent loop, tools, and TUI are unchanged across protocols.
• Reasoning display across all three protocols (StreamingEvent::ThinkingDelta): OpenAI reasoning_content (+reasoning alias / DeepSeek-reasoner), Gemini thought:true parts, Anthropic thinking_delta. Surfaced live (dim "💭 thinking"), kept out of the answer text and not echoed back next turn.
• 11 local-inference presets (OpenAI-protocol, keyless): ollama, lmstudio, vllm/llama.cpp/llama, mlx/omlx, llamafile, koboldcpp, localai, jan, oobabooga, gpt4all — each at its documented default port. --list-providers lists them.

TUI rendering

• Markdown italic, dependency-free code-fence syntax highlighting (syntax.rs), OSC-8 hyperlinks + bare-URL autolinking, word-level diff highlighting (word_diff.rs).
• Theme system (theme.rs): dark/light/high-contrast via --theme, covering syntax highlight + diff preview + markdown accents.
• Live spinner shows dynamic ↑/↓ token transfer (prompt vs output) plus a tok/s throughput pulse.
• CJK/wide-char display-width handling (width.rs); strip_ansi now also strips OSC sequences.

Agent KV-cache

• Rolling Anthropic prompt-cache breakpoint on the conversation tail (3 of 4 breakpoints: tools + system + conversation) so each turn reads the prior prefix from cache. Cached-token cost double-count fixed for OpenAI/Gemini.

Headless mode

• json/stream-json exit-code contract (non-zero on failure, envelope still printed); stream-json tool/thinking events; permission_request emission + pre-queued permission_response handling.

Verification

cargo build / test / clippy --all-targets / fmt --check all green — 1269 tests, 0 warnings.

Known follow-ups (need live-API e2e, out of scope here)

• Gemini request-side thinkingConfig enablement; Gemini cachedContent provisioning.
• Anthropic thinking-block + signature round-trip for multi-turn thinking+tools (pre-existing).
• Headless true bidirectional interactive permission (needs a background stdin-reader thread; this PR ships the pre-queue model).

🤖 Generated with Claude Code