feat(agent): implement tool-use orchestration loop

[Svagtlys]

Closes #79

Description

Implements the reason → act → observe tool-use orchestration loop: detect model tool calls, execute them against MCP servers, feed results back as provider-native tool messages, and re-invoke for final synthesis. Library-only primitive — no routes/lifespan wiring (Integration & Testing #1 composes it).

Implementation

• octave.inference types: ToolCall(id, name, arguments: str); Message gains tool_calls/tool_call_id/name and "tool" role; CompletionResult.tool_calls (additive, OpenAI-dialect)
• OpenAIAdapter: parses choice.message.tool_calls from completion responses; serializes tool-carrying messages to provider payloads with exclude_none (strict local engines reject null tool keys) and the function-call envelope
• octave.tools.errors: ToolError base introduced; ToolTranslationError re-based under it (backward-compatible)
• octave.agent (new package): composition layer and future Agent Manager home
  • ToolLoop — round mechanics, provider-invariant transcript ordering, max_tool_rounds limit raising ToolLoopLimitError with partial transcript; two-tier error split (Tier-1 model-correctable → Error: -prefixed tool messages; Tier-2 fatal → raise)
  • ToolExecutor Protocol — execution seam; implementations never raise for tool-level failures
  • McpToolExecutor — wraps ToolRegistry; joins text blocks, (no output) sentinel, McpError → ToolOutcome(is_error=True) conversion
  • SDK quarantine: AST guard (tests/agent/test_package.py) proves no openai/mcp imports; only mcp_executor.py touches octave.mcp
• docs/ARCHITECTURE.md: octave.agent section + MCP Connector schema-translation status updated

References

• Design doc: .agents/specs/2026-09-25-tool-use-orchestration-loop-design.md
• Implementation plan: .agents/specs/2026-09-25-tool-use-orchestration-loop.md

Checklist

• Tests passing (317/317)
• Lint clean (ruff check src tests)
• Types clean (mypy src, 54 files)
• Frontend builds: not applicable (no frontend changes)

Notes for reviewer

• Plan deviation: the plan's loop-limit test asserted a 5-message partial transcript, contradicting spec Testing item 10 (transcript must end on the unfulfilled assistant tool_calls message, N+1 completions). The test was corrected to match the spec, which the implementation satisfies.
• The style: commit swept in plans/2026-09-25-issue-concurrent-tool-execution.md (deferred concurrent-execution issue draft referenced by the design doc).