feat: added tools registry. 6 core tools.

[Blankeos]

What changed

• Added an OpenCode-style tool system with a registry + 6 core tools (glob, list, read, write, edit, bash) under src/tools/*.
• Wired tool definitions into LLM requests via aisdk and surfaced tool call lifecycle into the UI (tool rows + previews) (src/tools/aisdk_bridge.rs, src/ui/components/chat.rs, src/llm/client.rs).
• Implemented provider-aware system prompt composition (OpenAI/Anthropic/Gemini/Codex heuristics) and automatic rules loading from AGENTS.md / CLAUDE.md (local upward traversal + global fallbacks, size limits, disable env flags) (src/prompt/mod.rs, src/prompt/rules.rs).
• Added /refreshmodels to force-refresh the models.dev cache (src/command/handlers.rs, src/model/discovery.rs).
• Improved streaming metrics: persist timing primitives (t0_ms/t1_ms/tn_ms) + output_tokens, and display total latency + TTFT + TPS in chat metadata (src/session/types.rs, src/persistence/*, src/ui/components/chat.rs).
• Added planning docs under _plans/ that capture the intended tool/prompt architecture.

Why

• Enable first-class tool calling (file ops + bash) and better agent ergonomics.
• Make prompts consistent with OpenCode/Claude Code conventions via automatic rule discovery.
• Provide accurate, spec-aligned latency metrics (TTFT/TPS/total) instead of approximate “duration since first token”.
  Notes / reviewer callouts
• Cargo.toml patches aisdk to a local path (/Users/carlo/Desktop/Projects/aisdk-rs); this likely needs to be reverted to a git/registry dependency before merge.
• Tool-call parsing fallback exists (src/llm/tool_calls.rs) but primary execution is via aisdk tool streaming (src/tools/aisdk_bridge.rs).