feat(provider): add GigaChat via the gigachat library

[Rai220]

Summary

Adds GigaChat as a first-class LLM provider (model prefix gigachat::), routed through the official gigachat library rather than the OpenAI SDK — GigaChat is not OpenAI-compatible (OAuth, function role instead of tool, a single function_call instead of a tool_calls array).

Default model: gigachat::GigaChat-3-Ultra. Everything is env/Settings-configurable.

What's included

• ouroboros/llm.py — gigachat:: routing + a dedicated _chat_gigachat path (mirrors the existing _chat_anthropic) with native function-calling:
  • OpenAI tools → GigaChat functions; function_call: "auto"/"none"; native function_call response parsed back into OpenAI tool_calls.
  • Handles two GigaChat-specific quirks discovered during live testing: function results must be valid JSON (plain text is wrapped as {"result": ...}), and the system message must be first (later system-reminders are demoted to user).
• Auth — GIGACHAT_CREDENTIALS (authorization key, OAuth) or GIGACHAT_USER/GIGACHAT_PASSWORD (basic auth). GIGACHAT_SCOPE / GIGACHAT_BASE_URL / GIGACHAT_VERIFY_SSL_CERTS / GIGACHAT_PROFANITY_CHECK are all configurable.
• Provider wiring — config.py, pricing.py, safety.py, server_runtime.py, model_catalog_api.py (lists models via the library's native async client), agent_task_pipeline.py, launcher_bootstrap.py, contracts/plugin_api.py, server.py (secret redaction + restart keys), provider_models.py.
• Settings UI — GigaChat card (settings_ui.js/settings.js) + provider icon.
• Dependency — gigachat>=0.2.1 (lazy-imported).
• Tests — new tests/test_gigachat_conversion.py (converters: tool↔function, tool_calls↔function_call, JSON-result wrapping, system-demotion, response normalization) + provider-wiring updates across 8 test files.
• Docs — README, ARCHITECTURE (env table + routing), DEVELOPMENT, CHECKLISTS, CREATING_SKILLS.

Verification

• Unit + provider-routing tests pass; no regressions vs the pre-existing baseline.
• Verified live against a real GigaChat endpoint: direct chat, a tool-calling round-trip (get_weather), and a full agent task (read a file via tool → summarize) all succeed on GigaChat-3-Ultra. The model catalog lists real gigachat:: models.

Notes / known limitations

• GigaChat supports one function_call per turn; parallel OpenAI tool_calls are collapsed to the first.
• Cost is estimated from token counts (no live cost API).
• Not wired (no OpenAI-tools equivalent): few_shot_examples, return_parameters, function_ranker, forced-specific-function, functions_state_id threading.

🤖 Generated with Claude Code