v0.29.2 - prompt cache awareness

Highlights

• Improves API cost visibility by parsing provider prompt-cache usage, including DeepSeek cache hit/miss fields.
• Adds a stable OpenAI prompt_cache_key and cache-friendly prompt layout while preserving current-turn recency.
• Adds Target Context to every model evidence pack so the copilot model sees its read-only supervisor role, active scope, target session(s), and no-tools/no-injection boundary.

Verification

• Local: uv run --extra tui python -m unittest discover (728 tests OK)
• GitHub Actions: main tests, tag tests, and publish all passed