[AdaAibaby]

What

Add orchestrator.sandbox.checkpoint.change_footprint — a per-pause counter sliced by file_type (memfile|rootfs) and clean (dirty bytes <= one block), emitted from recordSnapshotDiff off data already computed on every pause. No behavior change.

Why

Agent sandboxes checkpoint constantly, and many turns change little. Whether to skip an unchanged checkpoint (or downgrade it to filesystem-only) is only worth building if the no-op rate is actually high — and today there is no metric for it. snapshot.diff.{bytes,ratio} exist but carry the full dirty size; nothing says "this pause was a no-op," and nothing is sliced so a dashboard can read the skip opportunity directly.

This counter is the measurement before the mechanism:

• per-dimension clean rate = the filesystem-only-downgrade / skip opportunity, measured on real traffic;
• both dimensions clean on one pause = the full-skip rate (join by sandbox/trace in the dashboard).

It needs no new machinery, no in-guest agent, no kernel features — just exposes the authoritative dirty-page count the orchestrator already derives every pause.

Context

A prior attempt to add skip-unchanged checkpoints (#2586, closed as "complex, benefits uncertain") bet on an in-guest eBPF inspector to manufacture a change signal. But the orchestrator already computes the exact dirty set per pause — the missing piece was never the signal, it was a way to quantify the opportunity so the feature can be decided on data rather than a guess. This PR is exactly that quantification, and nothing more.

Change

• meters.go: new counter name + description + unit.
• snapshot_metrics.go: register the counter; emit Add(1, {file_type, clean, team}) in recordSnapshotDiff; extract snapshotDiffClean(dirtyBytes) (threshold = one 4 KiB block) as a tested helper.
• unit test for the helper.

Caveat (documented separately)

On the destroy-path pause the memfile diff is parented on the original template, not the previous snapshot, so its "clean" reflects the resume working set, not since-last-snapshot change. The rootfs dimension is a true diff and is unaffected. Filed separately as a baseline-semantics clarification so the metric is not misread for memory.

[AdaAibaby]

Related issues surfaced while building this metric:

• snapshot.diff.ratio memfile baseline is path-dependent (template vs last-snapshot) — misleads skip/no-op analysis #3754 — snapshot.diff.ratio memfile baseline is path-dependent (the caveat above)
• Destroy-path memory diff not chained on previous checkpoint — re-stores full working set every pause #3755 — destroy-path memory diff not chained, re-stores full working set every pause