docs(memory-bank): record relay hardening policy baseline and operating knobs

memory-bank/decisions.md

@@ -45,3 +45,38 @@ Related:
 
 - `context/project-overview.md`
 - `memory-bank/README.md`
+
+---
+
+### D-002: Relay forwarding hardening policy baseline (P2-S3)
+
+date: 2026-03-24  
+author: Cline  
+scope: `pages/api/middleware/relayForwarding.js`, relay-backed endpoint handlers  
+type: decision  
+rationale: Consolidate relay reliability behavior into one shared policy so retries, timeouts, and logging are predictable and safe across all migrated GET flows.  
+impact: Improves resilience and observability while reducing risk of accidental overload, noisy duplicate logs, and inconsistent retry behavior per endpoint.  
+status: accepted
+
+Decision:
+
+- Shared relay defaults and bounds are centrally enforced in middleware:
+    - timeout default `8000ms`, clamped to `100..30000ms`
+    - retries default `2`, clamped to `0..4`
+    - retry delays clamped to `0..5000ms`
+- Retry classification is explicit:
+    - retryable: `408`, `429`, `5xx`, selected network/transient transport error codes
+    - non-retryable: deterministic client/auth statuses (`400`, `401`, `403`, `404`, `422`) and other explicit non-retryable HTTP statuses
+- Structured relay logs are redacted and emitted once per failure path:
+    - relay layer emits retry/failure structured events
+    - endpoint catch logging avoids duplicate emission when relay layer already logged terminal failure
+
+Consequences:
+
+- Endpoint contracts remain unchanged while reliability behavior is normalized.
+- Future relay policy changes should be made in one location and validated via phase21 hardening tests.
+
+Related:
+
+- `memory-bank/change-log.md` (CL-030, CL-031)
+- `tests/phase21/relay-forwarding-hardening.test.cjs`