Make MCP writes retry-safe and coordination state visible to agents

- Send Idempotency-Key on node create/patch/update writes (including bulk,
  link, move, kanban and agent-node tools) and return the key used, so an
  ambiguous failure can be retried without duplicating the write.
- Record a non-secret marker when login renewal is rejected; nap_whoami
  reports auth_required ahead of gateway state, and local-index answers
  carry stale warnings instead of silently serving old memory.
- Add nap_kanban_claim_next: pick and claim in one call, skipping cards
  held by another operation, never replaying an ambiguous claim.
- Add nap_changes_since over /v1/projects/{id}/changes with a
  per-worktree cursor that only advances after a successful read.
- Report real links_count/backlinks_count from indexed edges.
- Accept optional tool_profiles in .napseer/project.json when
  NAPSEER_TOOL_PROFILES is unset.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

tests/test_mcp_memory_surface_efficiency.py

@@ -36,7 +36,7 @@ def test_default_surface_includes_knowledge_files_and_coordination(mod):
     assert [tool["name"] for tool in mod.tools()] == [
         "nap_apropos", "nap_man", "nap_doctor", "nap_whoami",
         "nap_operation_list", "nap_operation", "nap_file_list", "nap_file_upload", "nap_file_download", "nap_file_archive",
-        "nap_discover", "nap_context", "nap_node_by_path", "nap_node_get",
+        "nap_changes_since", "nap_discover", "nap_context", "nap_node_by_path", "nap_node_get",
         "nap_create_node", "nap_node_patch", "nap_bulk", "nap_batch", "nap_ln",
         "nap_mv", "nap_rm",
     ]