Context-Aware Model Selection — selector suggest mode + logging

[DevVig]

Summary
Introduce a lightweight selector that recommends a model per task, with default static mapping preserved. Run in "suggest" mode by default and log rationale and final choice for auditability.

Why

• Optimize cost/performance without hidden complexity
• Keep choices reproducible and reviewable

Scope

• Add selector function and logging; keep static mapping as default
• Caps for latency/cost; do not introduce multi-turn loops

Tasks

• Implement a2a/models.py: select_model_with_reasoning(task, context) -> str (temp=0)
• Default to existing static map; add --selector-suggest flag to allow suggestion display
• Add logging: docs/timeline/model-selection.jsonl with {task, context_keys, recommended, chosen, rationale, ts}
• Config caps: max model/latency budgets in .a2dev/perf.yaml
• Unit tests: default path unaffected; suggest path logs without changing chosen model
• Update docs: configuration and usage

Acceptance Criteria

• No behavior change without flags
• Suggest mode logs recommendations and rationale
• All selections remain reproducible with static map unless explicitly overridden

Risks & Mitigations

• Hidden complexity -> keep suggestion-only default; explicit override required to change model
• Latency -> single call with short context keys, not full corpus

References

• suggestion.md (Context-Aware Model Selection)
• a2a/models.py