feat(adapters): add cohere-chat adapter for Cohere Chat API v2

[felmonon]

What

Adds a third provider adapter: cohere-chat, importing Cohere Chat API v2 conversation transcripts into the canonical trace format, with auto-detection.

Design notes

• Detection ordering matters: Cohere v2 deliberately shares the tool_calls / role: "tool" wire shape with OpenAI. Detection therefore scans the full message list for Cohere-only markers first — tool_plan on assistant messages, assistant message-level citations, and document typed tool-result blocks — before falling through to the per-message OpenAI/Anthropic rules. Covered by dedicated tests, including an ambiguous transcript that carries tool_calls but no tool_plan.
• tool_plan is agent behavior: it maps to a model_response step (marked metadata.source_field: "tool_plan") so plan text stays visible to safety scanning without being mistaken for the final answer.
• Document-block flattening: tool results arrive as {type: document, document: {data}} block lists; valid data payloads are joined in order, and malformed blocks are recorded under metadata.adapter.notes (never silently dropped) while still recovering the valid payloads around them.
• Cannot-carry fields stay honest: latency/cost/tokens/status are left unset so dependent checks go not applicable rather than silently passing — same contract as the existing adapters.

Testing

• 11 new tests: detection (3 incl. the shared-wire-shape trap), full refund-transcript mapping, tool_plan marking, citation mapping, edge-case note recording, round-trip validation, and a cross-provider verdict-equivalence test (same conversation via OpenAI and Cohere transcripts yields identical rule verdicts under the same policy).
• Full suite: 205 tests green on 3.11/3.13 locally; ruff and mypy --strict clean; benchmark suite thresholds all met.

Version

Bumps to 0.2.2 (new feature, backward compatible). Changelog updated.