docs(fleet): add operator configuration guide

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs/fleet/concepts/desired-vs-observed-state.md

@@ -0,0 +1,42 @@
+# Desired, Derived, and Observed Fleet State
+
+## One writable authority
+
+The canonical local v2 roster at `<MOSAIC_HOME>/fleet/roster.yaml` is desired state. Membership, stable identity, class, runtime/provider/model selection, launch policy, enablement, and persisted `running`/`stopped` intent are written only through generation-guarded roster mutations.
+
+Derived projections are reproducible consequences of that authority:
+
+- `<name>.env.generated`;
+- exact roster-named tmux sessions on the configured socket after reconciliation;
+- systemd service targets managed by installation/reconciliation.
+
+Current systemd unit enablement is not yet lifecycle-conformant at boot: installation can enable every
+agent unit, and the launcher projection does not carry `enabled` or `desired_state`. Therefore reboot
+preservation for stopped/disabled agents remains an FCM-M3-002 acceptance hold, not a guaranteed
+projection behavior.
+
+Observed evidence available to current roster-v2 status commands includes systemd active state, tmux
+presence, holder ownership, and unmanaged sessions. Heartbeat files are observational in the wider fleet,
+but roster-v2 `status`, `doctor`, and `verify` do not currently read them. Observation never writes back
+to the roster.
+
+## Generation and ownership
+
+`generation` is a positive integer concurrency fence. A mutating request must provide the current value. Successful changed CRUD increments it exactly once; stale or concurrent writers fail before mutation. Apply/reconcile rereads the canonical roster under a private exclusive lock and uses only that generation and content for effects.
+
+Ownership is exact, never fuzzy. Reconciliation is limited to roster names, the configured socket, the exact holder session, a private installation identity, and private managed paths. An ownership mismatch, unmanaged session, unsafe path, stale generation, or ambiguous lock fails closed.
+
+## Drift decisions
+
+| Observation                              | Interpretation           | Safe response                                                          |
+| ---------------------------------------- | ------------------------ | ---------------------------------------------------------------------- |
+| Generated file differs or is missing     | Derived projection drift | Review `apply --dry-run`; regenerate from the roster.                  |
+| Desired `running`, exact session missing | `missing-session`        | Diagnose ownership/runtime, then reconcile if safe.                    |
+| Desired `stopped`, exact session present | `unexpected-session`     | Inspect; reconciliation may stop only the proven roster target.        |
+| Disabled agent running                   | `disabled-running`       | Inspect; disabled state wins during explicit safe reconciliation.      |
+| Unknown session on the configured socket | Unmanaged state          | Report only. Do not adopt, rename, or kill it.                         |
+| Heartbeat stale in the wider fleet       | Liveness evidence        | Diagnose separately; current roster-v2 status does not read heartbeat. |
+
+`status` and `doctor` classify. `verify` is also observational but exits non-zero when ownership, drift, or unmanaged-state checks fail. `plan`/`apply --dry-run` validates proposed projection and lifecycle work without mutation. `apply` and `reconcile` converge only after all preconditions pass.
+
+A partial projection failure does not roll the roster back. Treat the committed roster as authority and regenerate. A lifecycle failure after projection completion preserves both roster and projections for inspection. Sensitive legacy values are never printed; diagnostics are bounded to stable codes, key names where applicable, and hashes.