chore: iterate dict directly instead of via .keys() (ruff SIM118)

[alexzhu0]

Summary

Ruff rule SIM118 flags x in dict.keys() / for x in dict.keys(). Iterating a dict directly is the idiomatic form and is measurably faster:

• .keys() constructs a fresh dict_keys view object on every call
• for x in d / x in d skips the view and goes straight to the dict's internal __iter__ / __contains__

5 sites, 4 files.

 # Step 2: Expand neighbors
 ids_1_set = set(relevant_ids)
-neighbor_ids = [nid for nid in fragment_1.nodes.keys() if nid not in ids_1_set]
+neighbor_ids = [nid for nid in fragment_1.nodes if nid not in ids_1_set]

Semantic equivalence

Every site either iterates the dict or checks membership — none captures, returns, or type-annotates the .keys() view. 4 of the 5 fixes were gated as "hidden" by ruff's conservative default because .keys() returns a distinct view type; I inspected each site manually and applied them via --unsafe-fixes after confirming no view is ever retained.

Files

File	Context
m_flow/retrieval/episodic/bundle_search.py	2 sites in _two_phase_projection / _two_phase_projection_with_stats — hot path for neighbor expansion
m_flow/retrieval/utils/procedural_bundle_search.py	same neighbor-expansion pattern in the procedural search variant
m_flow/adapters/vector/chromadb/ChromaDBAdapter.py	ChromaDB deserialization loop
m_flow/tests/test_multi_tenancy.py	a tiny _get_first_dataset_id test helper

4 files, +5 / −5.

Verification

$ ruff check . --exclude 'm_flow/baml_client' --exclude 'm_flow-frontend' --select SIM118
All checks passed!

$ for f in <each touched file>; do uv run python -c "import ast; ast.parse(open('$f').read())"; done
# clean across all 4 files

No behavior change. Generated code under m_flow/baml_client/ is untouched.

I affirm that all code in every commit of this pull request conforms to the terms of the M-flow Developer Certificate of Origin