refactor(api): tighten relay retry policy and config safety bounds

memory-bank/change-log.md

@@ -1148,3 +1148,44 @@ Validation:
 Follow-ups:
 
 - Optional next hardening increment: introduce endpoint-specific retry overrides for write paths (if future non-GET use is introduced) to keep retry policy conservative by operation type.
+
+---
+
+### CL-031: TASK22236 P2-S3 Batch 2A policy tightening (retry classification, config clamping, log de-dup)
+
+date: 2026-03-24
+author: Cline
+scope: `pages/api/middleware/relayForwarding.js`, `tests/phase21/relay-forwarding-hardening.test.cjs`
+type: change
+rationale: Execute P2-S3 Batch 2A by tightening relay retry policy and operational safety bounds while preventing duplicate error noise between relay-layer and endpoint-layer logging.
+impact: Stronger resilience and cleaner observability with no endpoint contract changes; retry behavior is now explicitly conservative for deterministic client/auth failures.
+status: completed
+
+Summary:
+
+- Updated relay retry policy:
+    - explicitly non-retryable statuses: `400`, `401`, `403`, `404`, `422`
+    - retries still allowed for transient classes (`408`, `429`, `5xx`) and selected transport error codes
+    - any other explicit numeric HTTP status now treated as non-retryable by default
+- Added runtime-safe config clamping for relay knobs:
+    - timeout clamped to `100..30000ms`
+    - retries clamped to `0..4`
+    - retry delays clamped to `0..5000ms`
+    - both env-derived and per-call numeric overrides are sanitized
+- Reduced duplicate logging noise:
+    - non-retry terminal relay failures are marked as already logged in middleware
+    - `relayGet` catch now skips `consoleLogger` when relay layer has already emitted a structured log
+- Expanded relay hardening tests:
+    - no retry on `401`
+    - duplicate logging suppression path through `relayGet`
+    - env-value clamping behavior for timeout bounds
+
+Validation:
+
+- `node tests/phase21/relay-forwarding-hardening.test.cjs` -> pass (6/6)
+- `node tests/phase21/endpoint-handler-contract.test.cjs` -> pass (152/152)
+- `npm run lint` -> warnings only (pre-existing `react-hooks/exhaustive-deps`; no new lint errors)
+
+Follow-ups:
+
+- Batch 2B: update memory/context docs with canonical relay hardening policy, env knobs, and rollback/tuning guidance.