shared-tree checkpoint: one pass covers every live peer and converges

[justrach]

Fixes #1622.

Root cause

presence.gateCheckIdentity acknowledged one peer per checkpoint, kept at most 16 acknowledgments, and never dropped a peer that had exited. With peers respawning, the set filled up. Every new peer after that could never be acknowledged, so every mutating call was refused for good, and each re-issue named a different peer.

Fix

• New presence_ack.ackLive. One checkpoint acknowledges every peer that is live at that moment, and drops acknowledgments of peers that are gone.
• An identical re-issue now proceeds. Only a peer that starts later fires another checkpoint.
• The checkpoint text names the first peer and how many others it covers. The NOT performed marker the batch gate matches is unchanged.

Not changed

Commands whose only effect is remote (over ssh) or outside the worktree are still checkpointed while peers are live. They now pass after one re-issue instead of never.

Tests

• New: two live peers are covered by one checkpoint, and 40 respawning peers each checkpoint once and then proceed. The ack set never fills.
• Presence and checkpoint unit tests: 130 pass.
• Tier-1 pre-push gate is green.