SIENTIAPDE-1975: align the gateway write path with the Laborious

- GW1: write commands (C2) carry only the server_id; ServerConfigResolver
  resolves the connection (command override, live pooled connection,
  connectors, opc-servers) with a 60 s cache and a 10 s negative cache.
- GW2: write_status keeps the worst outcome (ok < error < session_bad,
  pending/NULL replaced), same rule as the Laborious, built with SQLAlchemy
  Core (no SQL formatting).
- GW3: update_write_status never raises; outcome (updated, kept,
  row_not_found, invalid, error) logged and counted
  (opc_write_status_update_total); runs off the event loop.
- GW4: the reason of a failed write is logged; opc_write_status_by_model
  shows the last outcome; a bad timestamp is a failed write, not an
  exception.
- GW5: the NATS first connect gives up after 5 s (nats-py's defaults block
  the main loop for about 4 minutes while NATS is down); reconnection after
  it is unbounded; idle_sweep restarts a connection closed for good; stop()
  is safe on a closed connection.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

README.md

@@ -56,7 +56,8 @@ Gateway for on-demand and command-driven interactions with OPC UA servers:
 
 - **NATS write consumer**: consumes write commands from `opc.write` (see
   `NATS_WRITE_SUBJECT`) and executes them against the target OPC server
-  (`ingestor/adapters/nats_consumer.py`, `ingestor/engines/write.py`).
+  (`ingestor/adapters/nats_consumer.py`, `ingestor/engines/write.py`). See
+  [Write commands](#write-commands-c2-and-write_status-c3) below.
 - **On-demand gRPC `ReadTag`**: a gRPC server (`ingestor/adapters/grpc_server.py`,
   `ingestor/proto/opc_gateway.proto`) for synchronous, single-tag reads.
 - **Scheduled active reads**: a polling read engine (`ingestor/engines/active_read.py`)
@@ -70,6 +71,36 @@ See `docs/deploy-notes.md` for the deployment-facing details (env vars,
 health probe, NetworkPolicy, cert cache) and the `# OPC Gateway` block in
 `.env.example` for the full list of new environment variables.
 
+### Write commands (C2) and `write_status` (C3)
+
+The Laborious publishes one command per tag (SIENTIAPDE-1975):
+
+```json
+{"server_id": "1", "addr": "ns=2;s=Pred", "value": 1.25, "data_type": "float",
+ "model_id": "2", "model_name": "sin-approx", "schema": "sientia_data",
+ "table_name": "predictions", "timestamp": "2026-10-13T12:00:00-03:00",
+ "write_timestamp": false, "write_quality": false}
+```
+
+- **Server connection**: the command carries only `server_id`. The connection is
+  resolved by `ServerConfigResolver` (`ingestor/connection/servers.py`): `url`/
+  `server_uri`/`use_security` in the command when present (override), else the
+  live pooled connection opened by the reads, else the `connectors` document
+  (`type: opc_ua`, `status: Active`), else the `opc-servers` document
+  (`use_security = bool(cert_path)`) — the orchestrator's rule for the read slots.
+  Lookups are cached for 60 s. An unknown server is a failed write (`server_not_found`).
+- **Outcome**: `UPDATE ... SET write_status` on the prediction row (by `timestamp` +
+  `model_id`), keeping the worst outcome: `ok` < `error` < `session_bad`, and
+  `pending`/NULL are replaced by any answer. The Laborious applies the same rule to
+  the outcomes it records itself, so the order of the answers does not matter.
+  The UPDATE never raises; its outcome (`updated`, `kept`, `row_not_found`,
+  `invalid`, `error`) is logged and counted in `opc_write_status_update_total`.
+  The reason of a failed write is logged by `WriteEngine`.
+- **NATS connection**: the first connection gives up after 5 s (it runs on the main
+  loop: nats-py's defaults would block it for minutes while NATS is down); once
+  connected, it reconnects forever. A connection closed for good is started again
+  by `GatewayRuntime.idle_sweep`.
+
 ## 📋 Prerequisites
 
 - Python 3.11+
@@ -290,6 +321,9 @@ The OPC Ingestor exposes comprehensive Prometheus metrics for monitoring:
 ### OPC Gateway Metrics
 - `opc_write_total`: Active write attempts (NATS-driven), by result
 - `opc_write_latency_seconds`: Active write latency
+- `opc_write_status_by_model`: Last write outcome per model (1 = ok, 0 = failed; label `status`)
+- `opc_write_status_update_total`: `write_status` UPDATEs, by result
+  (`updated`, `kept`, `row_not_found`, `invalid`, `error`)
 - `opc_active_read_total`: Scheduled/on-demand read attempts, by result
 - `opc_error_by_type_total`: OPC errors by type (connection, conversion, session, ...)