2026-04-29 - Phi Dead Artifact Cleanup Approval Intent

Intent: Record the user’s explicit approval to delete strong dead Phi artifacts and reclaim disk while preserving small tracked manifests where possible; this follows measurement discipline by removing only paths already ruled out by residency or slower-profile evidence.

Setup: Approved targets are the rejected phi4_mini_ane_24layer_probe [0,24) CPU-fallback artifact; generated top-k LM-head artifacts that failed ANE residency because ios18.topk/ios18.cast lowered to CPU; generated 3-way and 8-way LM-head artifacts that profiled slower than the 4-way head; and known slower 5-layer tail artifacts [20,25), [25,30), plus duplicate [30,32).

Result: Cleanup approval logged; deletion itself is a separate destructive action and should be limited to the approved dead artifacts.

Surprise / hurdle: Disk pressure forced distinguishing failed/slower experiment artifacts from the current working baseline instead of doing broad cleanup.

Lesson: Artifact cleanup is safe only when each deletion target has a recorded rejection reason and current baselines are explicitly protected.

Next: Delete only the approved dead artifacts if cleanup proceeds; do not delete the current baseline artifacts or any batch-4 LM-head artifacts.

Refs: research/ANE_CHAIN_SCHEMA.md