Add interactive chat REPL with persistent sessions

[rejojer]

Summary

Introduces openkb chat, a multi-turn conversation REPL for the knowledge base. Conversations are automatically persisted to .openkb/chats/<id>.json and can be resumed across invocations. Also bundles a handful of adjacent cleanups that were surfaced while building it.

Supersedes #16 (same content, squashed into logical commits, plus one new commit sanitizing image payloads from persisted chat history).

Main feature

• New openkb chat CLI command with --resume [ID] (latest or by id / unique prefix), --list, --delete ID, and --no-color.
• Built on prompt_toolkit: bottom toolbar, styled prompt, Ctrl-C aborts current response while Ctrl-D / /exit quit cleanly.
• In-REPL slash commands: /exit, /clear (start a fresh session, previous is kept on disk), /save [name] (export markdown transcript to wiki/explorations/), /help.
• Streams the agent response with the same visual language as openkb query (dim tool-call lines flush against the preceding text, blank line separating a tool batch from the following response paragraph).
• Header mirrors the Claude Code layout: brand-colored title with version, then <kb dir> · <model> · session <id>, then a short key hint.
• Resume view shows prior turns compactly (older turns collapsed to ... N earlier turns omitted).
• Reuses the existing build_query_agent so tool behavior is identical to single-shot query.
• Persisted history strips base64 image payloads from get_image tool outputs and replaces them with a text reference (source path + "call get_image again" hint), so .openkb/chats/<id>.json stays small and legacy sessions are sanitized on load.

Prompt / agent polish bundled in

• Fix a copy-paste bug where the Q&A and lint agents were told to "Write all wiki content in X language" — they don't write wiki content. Switched to "Answer in X" and "Write the lint report in X".
• Give all three agents an "OpenKB" identity so the model introduces itself consistently.
• Finish the Q&A search-strategy step on summaries (tell the model to follow full_text when a summary is too thin), move the get_image "when to call" guidance into the tool docstring, and reword step 5 to refer to the tool by name.

Refactor bundled in

• Normalize get_page_content naming: rename the helper in openkb.agent.tools to get_wiki_page_content (matching the *_wiki_* sibling convention), rename the agent-side wrapper from get_page_content_tool to get_page_content so the tool name the model sees matches what the instructions have always said, drop the lazy-import workaround, and fix a stale test assertion.
• Derive openkb.__version__ from installed package metadata (importlib.metadata) so pyproject.toml is the single source of truth for the version string.

Test plan

• openkb chat from a kb root starts a new session, header renders, prompt works
• Single-turn question streams a response, tool-call lines render dim, blank-line spacing is correct
• Multi-turn: follow-up questions ("what about that?") use conversation context
• /save writes a readable transcript to wiki/explorations/
• /clear starts a fresh session; previous shows up in openkb chat --list
• Ctrl-C during streaming aborts the current response but keeps the REPL alive
• Ctrl-D / /exit quit with the friendly goodbye message
• openkb chat --resume picks up the latest session; --resume <prefix> resolves a unique prefix
• openkb chat --list and --delete <id> work
• After a turn that calls get_image, the saved .openkb/chats/<id>.json no longer contains the base64 data URL
• Existing openkb query still works (shared agent)
• pytest tests/test_agent_tools.py tests/test_query.py tests/test_chat_session.py passes