Improve default compaction prompt for continuation-state fidelity

### What do you want to change? Would you be open to adjusting the default compaction prompts so repeated summaries merge, deduplicate, and reconcile state rather than mainly preserving prose? For coding sessions, I would like the checkpoint to distinguish directly observed results from assistant-reported results, preserve continuation-critical literals such as paths, IDs, commands, hashes, and errors, and avoid reviving completed or superseded TODOs. ### Why? This extends the concern in #6157. In a small local single-session experiment with a custom `/cleanup` extension, the earlier concise prompt retained 73.9% of annotated continuation-critical literals. A state-ledger prompt across three runs retained 91.3–100% and all annotated open items. The main remaining failure was evidence calibration: documentation or static code could still be mistaken for runtime verification. This is not a general benchmark and does not describe Pi's default prompt results; it only suggests that prompt structure materially affects continuation quality. ### How? (optional) A small prompt-only experiment could add merge/dedup/reconcile rules, separate evidence status from task lifecycle, require exact engineering literals, and run a final coverage check. I can share the sanitized prompt and test notes if useful. I would appreciate maintainer feedback before attempting a change.