[BUG]: Prune clears read results causing instruction re-attachment; prune early-exit skips older prunable tools #30807

Description Two bugs in the pruning mechanism. Bug 1: Prune causes instruction file re-attachment instruction.ts:20 鈥?when scanning previously read file paths for auto-attaching instruction files (AGENTS.md, CLAUDE.md), if part.state.time.compacted is set, the read result is skipped. After prune() clears old tool output, all read results that were pruned are invisible to this scan. The system re-attaches instruction files that were already loaded in the current session, causing duplication or re-sending instructions that were already applied. Bug 2: Prune early-exit breaks loop compaction.ts:325 鈥?if (part.state.time.compacted) break loop. This breaks out of the entire message loop, not just the current message's parts. If a more recent message has a compacted tool from a previous prune run, pruning stops entirely, leaving older prunable tools uncleaned. Focused follow-on to #30376. Suggested fix: (Bug 1) preserve read-file metadata separately from tool output so it survives pruning. (Bug 2) change break to continue so the loop checks the next message. ### OpenCode version

main

Steps to reproduce 1. Session reads AGENTS.md via Read tool (turn 5) 2. After many turns, prune() clears older tool output 3. instruction.ts re-attaches AGENTS.md in next turn (turn 50+) even though it was already loaded 4. AI receives duplicate instructions