llm-infra scan: litellm-0001 _delete_deployment combined_id_list O(R*C); haystack CLEAN

litellm-0001: ProxyConfig._delete_deployment() checks model_id not in
combined_id_list (a plain Python list) for each router model ID during
config refresh. O(R*C) where R=router models, C=combined list. Fix:
convert to set before loop. MEDIUM severity, 6.5x at N=500. 1/1 PASS.

Haystack: no CWE-407 defects found. Uses dicts and sets throughout
for membership checks (scheduled_components set, storage dict, etc.).

defects/haystack/CLEAN.md

@@ -0,0 +1,18 @@
+# Haystack - CWE-407 Scan Result: CLEAN
+
+Scanned: 2026-03-30
+Target: https://github.com/deepset-ai/haystack (depth=1)
+Focus: pipeline component dedup, document store dedup, retriever result dedup, node connection membership
+
+## Findings
+
+No CWE-407 defects found. The codebase uses appropriate data structures throughout:
+
+- Pipeline scheduler uses `set()` for `scheduled_components` (async_pipeline.py:214)
+- Breakpoint validation uses `set()` for component validation (breakpoint.py:97,100)
+- Document store uses `dict` for ID lookups (in_memory/document_store.py)
+- Embedding dedup uses `dict` for `doc_ids_to_embeddings`
+- Graph nodes/edges use NetworkX dict-based adjacency
+- `model_names` equivalent lookups use dict views
+
+The deepset team has made good data structure choices. No linear-scan-in-loop patterns found on any hot path.