Fixes the 'live peer looks disconnected' class of bugs. Two layers:
ROOT CAUSE — involuntary mesh context loss:
The session→mesh binding lived only in the daemon's in-memory registry,
so a daemon restart (e.g. `daemon down && up`) wiped it. Every live
session then lost its mesh, and CLI commands fell back to an arbitrary
default mesh — a peer that never moved looked offline.
Fix: persist session bindings to ~/.claudemesh/daemon/sessions.json
(secret-free — keypairs reload from the per-session keypair store). On
boot the daemon rehydrates each binding whose pid is still alive (with a
start-time PID-reuse guard), reloads its keypair, re-signs a parent
attestation, and re-registers it — which reconnects its SessionBroker
WS. Restarts are now transparent; sessions keep their mesh.
DEFENSIVE LAYER — cross-mesh send resolution:
`send` without --mesh and several joined meshes returned mesh_required;
a prefix under --mesh X resolved against the default mesh's roster, not
X's (only the full 64-char pubkey worked). Now a name/prefix is resolved
across all joined meshes (or scoped to --mesh): unique match auto-selects
its mesh, multi-mesh match asks for --mesh, none gives a clear error.
Kills mesh_required for peers on a non-default mesh and fixes P3.
Maps to field-report P1/P2/P3. P4 (shared member) left as-is (by design).
New: 5 persistence unit tests. Full suite 119/119. Daemon boot verified.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
c747040e0d