chore(deps-dev): update ruff requirement from <0.16,>=0.15.22 to >=0.15.22,<0.17

AGENTS.md

@@ -42,7 +42,7 @@ python -m pytest tests/ -q                                          # unit tests
 python -m eval.harness --dataset eval/datasets/sample.jsonl --k 5   # retrieval eval gate
 python -m eval.harness --dataset eval/datasets/codemem.jsonl --k 5  # larger eval; covers conflict resolution
 python -m eval.ablation                                             # vector-only vs 1-hop vs PPR
-ruff check .                                                        # lint (line-length 100, py39)
+ruff check .                                    # lint (line-length 100, py39, pinned rule set)
 
 # ── External benchmarks (real numbers need torch + the dataset; see eval/external.py) ──
 python -m eval.external --dataset locomo10.json --format locomo --k 10        # LoCoMo

pyproject.toml

@@ -126,13 +126,13 @@ all = [
     "onnxruntime<1.24; python_version < '3.11'",
     "psycopg[binary]>=3.1",
 ]
-dev = ["pytest>=8.0", "pytest-asyncio>=0.23", "ruff>=0.15.22,<0.16"]
+dev = ["pytest>=8.0", "pytest-asyncio>=0.23", "ruff>=0.15.22,<0.17"]
 # Everything needed to run the FULL offline gate in CI (lint + all extras-gated tests)
 # WITHOUT pulling torch/sentence-transformers — no test needs the real embedder.
 test = [
     "pytest>=8.0",
     "pytest-asyncio>=0.23",
-    "ruff>=0.15.22,<0.16",
+    "ruff>=0.15.22,<0.17",
     "python-dotenv>=1.0",
     "uvicorn[standard]>=0.29",
     "fastapi>=0.133.1,<1; python_version >= '3.10'",
@@ -195,6 +195,13 @@ include = ["engraphis*", "scripts*"]
 line-length = 100
 target-version = "py39"
 
+# Pin the rule set explicitly instead of inheriting ruff's built-in default, which is not
+# stable across minor releases -- 0.16 widened it from 59 rules to 413 and turned every
+# routine linter bump into a repo-wide breakage. These are exactly the 59 rules ruff
+# 0.15.x enabled by default, so behaviour is unchanged; widening is now a deliberate edit.
+[tool.ruff.lint]
+select = ["E4", "E7", "E9", "F"]
+
 [tool.pytest.ini_options]
 testpaths = ["tests"]
 addopts = "-q"

tests/test_release_infrastructure.py

@@ -115,7 +115,12 @@ def test_codeql_workflow_fails_when_sarif_contains_findings():
 def test_ci_linter_is_bounded_to_the_verified_release_series():
     pyproject = _text("pyproject.toml")
 
-    assert pyproject.count('"ruff>=0.15.22,<0.16"') == 2
+    # A version bound alone never made the linter deterministic: ruff's *default* rule set
+    # is not stable across minor releases -- 0.16 widened it from 59 rules to 413, which
+    # would have turned `ruff check .` red on unchanged code. Pinning `select` explicitly
+    # is what actually bounds CI, so the bound and the rule set are asserted together.
+    assert pyproject.count('"ruff>=0.15.22,<0.17"') == 2
+    assert 'select = ["E4", "E7", "E9", "F"]' in pyproject
 
 
 def test_release_repair_requires_tag_sha_successful_build_publish_and_pypi_identity():