rho > micro_threshold: replace tool results with a compact excerpt. Selection via
[excerptable_tool_result_indices]; this executor only applies the excerpt.
Pure selection (W1-1 collapse): how many of the oldest history messages to drop to bring the
partition under target_tokens, never crossing the preserve-recent floor (keep messages).
Returns (count, tokens_saved); the executor just drains count from the front. This is the
decision the cache-aware planner reuses to “batch one big drop to target” rather than re-deriving
the count inside the compactor.