feat: per-model cost tracking and message attribution

[jmars]

Problem

AgentStats.session_cost uses a single flat pricing pair for the entire session. When /model switches the active model mid-session, all accumulated tokens are retroactively re-priced — update_pricing() even documents this as "a known approximation." Additionally, session transcripts (messages.jsonl) have no way to identify which model generated each message.

Solution

Two changes:

1. Per-model cost tracking in AgentStats

• New PerModelTokenUsage model tracks prompt_tokens, completion_tokens, and per-model pricing
• AgentStats.per_model: dict[str, PerModelTokenUsage] accumulates usage keyed by model name
• session_cost sums across all models at their respective rates (falls back to legacy single-model calc when per_model is empty)
• add_tokens_for_model() attributes each turn's usage to the active model; _ensure_per_model() snapshots the pricing at first attribution so switching A→B→A preserves correct rates
• update_pricing() now only sets the current-model default — past tokens are untouched

2. Model attribution on LLMMessage

• New optional model and provider fields (default None, omitted via exclude_none in serialization)
• Set on assistant messages in both streaming (_chat_streaming) and non-streaming (_complete) paths
• Propagated through chunk accumulation (LLMMessage.__add__) and model_dump

Before / after (cost example)

Session: 1M prompt + 50K completion on deepseek ($0.14/$0.28/M)
Switch to claude ($3.00/$15.00/M), one turn: 1.5M prompt + 30K completion

Before:  $1.45  (all tokens at claude pricing — wrong)
After:   $0.15 + $4.95 = $5.10  (each model at its own rate — correct)

Backward compatibility

• Old messages.jsonl (no model/provider): fields default to None, serialization omits them
• Old meta.json (no per_model): defaults to empty dict, session_cost falls back to legacy calculation
• No migration needed — per_model populates lazily on next turn

Token attribution note

Each turn reports the full context prompt_tokens the API billed for, not just the incremental delta. When a model switch occurs, the new model's turn includes the cost of re-encoding the entire history — so per-model totals correctly reflect what each model was actually charged. Prompt caching discounts and tokenization differences are provider-side and not reflected in usage responses.

Files changed

• vibe/core/types.py: +80/-8 (PerModelTokenUsage, AgentStats.per_model + methods, LLMMessage.model/provider)
• vibe/core/agent_loop/_loop.py: +24/-2 (message stamping, model_name param on _update_stats)