Add fixture specialization detectors

Author: SinatrasC

.github/workflows/blue-pr-prechecks.yml

@@ -1,24 +1,20 @@
 name: blue-pr-prechecks
 
 on:
-  pull_request:
-    types:
-      - opened
-      - synchronize
-      - reopened
-      - ready_for_review
-      - edited
+  workflow_dispatch:
 
 permissions:
   contents: read
   pull-requests: read
 
 concurrency:
-  group: blue-pr-prechecks-${{ github.event.pull_request.number }}
+  group: blue-pr-prechecks-${{ github.event.pull_request.number || github.run_id }}
   cancel-in-progress: true
 
 jobs:
   prechecks:
+    # Temporarily disabled while blue submissions are inactive.
+    if: ${{ false }}
     runs-on: ubuntu-latest
     timeout-minutes: 20
     env:

.github/workflows/blue-pr-sync.yml

@@ -1,23 +1,20 @@
 name: blue-pr-sync
 
 on:
-  pull_request_target:
-    types:
-      - opened
-      - synchronize
-      - reopened
-      - edited
+  workflow_dispatch:
 
 permissions:
   contents: read
   pull-requests: read
 
 concurrency:
-  group: blue-pr-sync-${{ github.event.pull_request.number }}
+  group: blue-pr-sync-${{ github.event.pull_request.number || github.run_id }}
   cancel-in-progress: true
 
 jobs:
   gate:
+    # Temporarily disabled while blue submissions are inactive.
+    if: ${{ false }}
     runs-on: ubuntu-latest
     timeout-minutes: 10
     outputs:

.github/workflows/trusted-blue-eval.yml

@@ -12,35 +12,22 @@ on:
         required: false
         default: true
         type: boolean
-  # Auto-fire after blue-pr-sync registers a new patch in the API queue.
-  # The claim step returns HTTP 204 when the queue is empty, so spurious
-  # triggers are no-ops.
-  workflow_run:
-    workflows: ["blue-pr-sync"]
-    types: [completed]
 
 permissions:
   contents: read
 
 env:
   KERNELGUARD_API_BASE_URL: https://kguard.sinatras.dev
 
-# Per-trigger concurrency group so multiple PRs syncing within seconds of
-# each other do not cancel each other out (GitHub allows only 1 pending +
-# 1 running per group, beyond which new runs auto-cancel). Each
-# blue-pr-sync run carries the candidate head SHA on
-# event.workflow_run.head_sha; manual workflow_dispatch falls back to the
-# unique run id so manual drains never collide with each other either.
+# Manual workflow_dispatch falls back to the unique run id.
 concurrency:
   group: trusted-blue-eval-${{ github.event.workflow_run.head_sha || github.run_id }}
   cancel-in-progress: false
 
 jobs:
   claim:
-    # Skip when triggered by a failed blue-pr-sync (rejected gate, static
-    # security failure, etc.). Manual workflow_dispatch always passes this
-    # check because workflow_run context is empty in that case.
-    if: ${{ github.event_name != 'workflow_run' || github.event.workflow_run.conclusion == 'success' }}
+    # Temporarily disabled while blue submissions are inactive.
+    if: ${{ false }}
     runs-on: ubuntu-latest
     timeout-minutes: 10
     environment: trusted-blue-eval

kernelguard.py

@@ -2846,6 +2846,149 @@ def detect_hardcoded_shapes(code: str | SubmissionFacts) -> list[dict]:
     return matches
 
 
+def detect_benchmark_unsafe_algo_dispatch(code: str | SubmissionFacts) -> list[dict]:
+    """Shape/data-family dispatch into unsafe specialized algorithms.
+
+    Per-shape tuning is legitimate. This detector requires the stronger
+    combination: benchmark-shaped branches plus data-family classifiers and
+    fast-path terms that imply reduced precision, truncated work, or a generic
+    exact fallback for cases outside the public fixture.
+    """
+    facts = ensure_submission_facts(code)
+    raw_code = facts.raw_code
+    active_code = facts.python_active
+    combined = f"{active_code}\n{raw_code}"
+    low = combined.lower()
+    matches = []
+
+    shape_gate_count = len(re.findall(
+        r'(?:\(\s*[Bb]\s*,\s*[Nn]\s*\)\s*==\s*\(\s*\d{1,5}\s*,\s*\d{1,5}\s*\)|'
+        r'\b(?:B|b|batch|N|n)\s*(?:==|!=|<=|>=|<|>)\s*\d{1,5})',
+        combined,
+    ))
+    benchmark_constant_count = len(re.findall(
+        r'\b(?:60|176|352|512|640|768|1024|2048|4096|1280)\b',
+        combined,
+    ))
+    family_terms = re.findall(
+        r'\b(?:classify_512|classify_1024|detect_label|rankdef|rank_def|clustered|'
+        r'nearrank|near_rank|nearcol|near_col|mixed|rowscale|colnorm|zerofrac|'
+        r'colrange|tail_max|neardiff|inactive_tail)\b',
+        low,
+    )
+    unsafe_terms = re.findall(
+        r'\b(?:zero_tail|trunc|truncate|stop_at|switch_at|ncap|ncols|limit|'
+        r'_trunc_stop|tf32|fp16|bf16|fast_path|unsafe|rankdef)\b',
+        low,
+    )
+    fallback_terms = re.findall(
+        r'\b(?:torch\.geqrf|torch\.linalg\.(?:qr|lu)|fallback|fall back|exact|generic)\b',
+        combined,
+        re.IGNORECASE,
+    )
+    raw_native_cache = (
+        bool(re.search(r'\bnvcc\b', low))
+        and bool(re.search(r'\b(?:ctypes|CDLL|subprocess\.run|Popen|\.so|\.cu)\b', combined))
+        and not bool(re.search(r'\bload_inline\s*\(', active_code))
+        and len(re.findall(r'\b(?:data_ptr|ctypes\.c_void_p|uint64_t|void\s*\*)\b', combined)) >= 4
+    )
+
+    strong_shape_fixture = shape_gate_count >= 8 and benchmark_constant_count >= 20
+    data_family_routing = len(family_terms) >= 4 and len(unsafe_terms) >= 4
+    exact_escape_hatch = len(fallback_terms) >= 2
+
+    if raw_native_cache and strong_shape_fixture and data_family_routing:
+        signals = [
+            "raw native compile/load cache",
+            f"{shape_gate_count} shape gates",
+            f"{len(set(family_terms))} data-family terms",
+            f"{len(set(unsafe_terms))} unsafe/truncation terms",
+        ]
+        if exact_escape_hatch:
+            signals.append("exact/generic fallback path")
+        matches.append({
+            "pattern": "BENCHMARK_UNSAFE_ALGO_DISPATCH",
+            "severity": "high",
+            "evidence": "; ".join(signals),
+        })
+
+    return matches
+
+
+def detect_input_structure_truncation_dispatch(code: str | SubmissionFacts) -> list[dict]:
+    """Input probes routing to partial computation or exact subset repair.
+
+    This is not a ban on algorithm selection. It targets classifiers/probes of
+    input structure that steer only a detected subset into truncated fast paths,
+    tail-zeroing, or reference/exact repair, which is the static shape of a
+    public-fixture assumption.
+    """
+    facts = ensure_submission_facts(code)
+    active_code = facts.python_active
+    raw_code = facts.raw_code
+    combined = f"{active_code}\n{raw_code}"
+    low = combined.lower()
+    matches = []
+
+    probe_terms = re.findall(
+        r'\b(?:probe_cond|probe_n|_probe|_probe_conditioning|_s3_collinear|'
+        r'spread_thresh|sparse_thresh|corr_thresh|bit0|bit1|classify_512|'
+        r'classify_1024|detect_label|colnorm|neardiff|zerofrac|tail_max)\b',
+        low,
+    )
+    subset_terms = re.findall(
+        r'\b(?:bad_idx|good_idx|bad_count|good_count|index_select|index_copy|'
+        r'scatter_exact|compact_label|panel_labels|label_bad_cache|gather_.*bad|'
+        r'compact_.*bad)\b',
+        low,
+    )
+    partial_terms = re.findall(
+        r'(?:tau\s*\[[^\]]*limit\s*:\s*\]\s*=\s*0|H\s*\[[^\]]*limit\s*:\s*\]\s*=\s*0|'
+        r'\b(?:zero_tail|trunc|truncate|inactive_tail|stop_at|switch_at|ncap|ncols|limit\s*<\s*n)\b)',
+        combined,
+        re.IGNORECASE,
+    )
+    repair_terms = re.findall(
+        r'\b(?:repair|exact|torch\.geqrf|torch\.linalg\.(?:qr|lu)|index_copy|index_select)\b',
+        combined,
+        re.IGNORECASE,
+    )
+    probe_set = set(probe_terms)
+    subset_set = set(subset_terms)
+    partial_set = set(partial_terms)
+    repair_set = set(repair_terms)
+
+    probe_driven_repair = (
+        bool(probe_set & {"probe_cond", "probe_n", "_probe", "_s3_collinear"})
+        and len(probe_set) >= 4
+        and len(subset_set) >= 3
+        and len(repair_set) >= 3
+    )
+    classifier_driven_partial = (
+        bool(probe_set & {"classify_512", "classify_1024", "detect_label"})
+        and len(probe_set) >= 4
+        and len(partial_set) >= 3
+        and len(subset_set) >= 2
+    )
+
+    if probe_driven_repair or classifier_driven_partial:
+        signals = [
+            f"{len(probe_set)} structure/probe terms",
+            f"{len(subset_set)} subset-routing terms",
+        ]
+        if partial_set:
+            signals.append(f"{len(partial_set)} partial/truncation terms")
+        if repair_set:
+            signals.append(f"{len(repair_set)} exact-repair terms")
+        matches.append({
+            "pattern": "INPUT_STRUCTURE_TRUNCATION_DISPATCH",
+            "severity": "high",
+            "evidence": "; ".join(signals),
+        })
+
+    return matches
+
+
 def detect_unsync_multistream(code: str | SubmissionFacts) -> list[dict]:
     """Pattern 14: Unsynchronized multi-stream dispatch to exploit timing.
 
@@ -7289,6 +7432,14 @@ class RulePolicy:
         "HARDCODED_SHAPES", "performance_heuristic", "telemetry", TELEMETRY_ONLY, (),
         (), "downgrade",
     ),
+    "BENCHMARK_UNSAFE_ALGO_DISPATCH": RulePolicy(
+        "BENCHMARK_UNSAFE_ALGO_DISPATCH", "fixture_specialization", "hard", AUTO_FILTER, (),
+        (), "keep",
+    ),
+    "INPUT_STRUCTURE_TRUNCATION_DISPATCH": RulePolicy(
+        "INPUT_STRUCTURE_TRUNCATION_DISPATCH", "fixture_specialization", "hard", AUTO_FILTER, (),
+        (), "keep",
+    ),
     "UNSYNC_MULTISTREAM": RulePolicy(
         "UNSYNC_MULTISTREAM", "timing_manipulation", "telemetry", TELEMETRY_ONLY, (),
         (), "downgrade",
@@ -7556,6 +7707,8 @@ def support_only_patterns(matched_patterns: list[dict]) -> bool:
     detect_torch_compile_cache,
     detect_cuda_graph_python,
     detect_hardcoded_shapes,
+    detect_benchmark_unsafe_algo_dispatch,
+    detect_input_structure_truncation_dispatch,
     detect_unsync_multistream,
     detect_cuda_event_disable_timing,
     detect_token_paste_cuda_api,
@@ -7617,6 +7770,8 @@ def support_only_patterns(matched_patterns: list[dict]) -> bool:
     ("torch_compile_cache", detect_torch_compile_cache),
     ("cuda_graph_python", detect_cuda_graph_python),
     ("hardcoded_shapes", detect_hardcoded_shapes),
+    ("benchmark_unsafe_algo_dispatch", detect_benchmark_unsafe_algo_dispatch),
+    ("input_structure_truncation_dispatch", detect_input_structure_truncation_dispatch),
     ("unsync_multistream", detect_unsync_multistream),
     ("cuda_event_disable_timing", detect_cuda_event_disable_timing),
     ("token_paste_cuda_api", detect_token_paste_cuda_api),
@@ -8522,7 +8677,7 @@ def _worker_parquet(args: tuple) -> dict:
     "RUNNER_PLAN_CACHE", "CUDA_GRAPH_PYTHON", "CUDA_GRAPH_REPLAY",
     "TIMER_MONKEYPATCH", "FAKE_BENCHMARK_EMIT", "STDIO_REDIRECT", "UNSYNC_MULTISTREAM", "CUDA_EVENT_DISABLE_TIMING",
     "SCALED_MM_REF", "DECODE_MM_REF", "SILENT_FALLBACK", "REFERENCE_PRECOMPUTE_REPLAY", "TORCH_COMPILE_CACHE",
-    "HARDCODED_SHAPES", "TRIVIAL_PROBE",
+    "HARDCODED_SHAPES", "BENCHMARK_UNSAFE_ALGO_DISPATCH", "INPUT_STRUCTURE_TRUNCATION_DISPATCH", "TRIVIAL_PROBE",
     "OBFUSCATED_EXEC", "DYNAMIC_EXECUTION", "MODULE_RELOAD", "THREAD_INJECTION", "LAZY_TENSOR",
     "TOKEN_PASTE_CUDA_API", "SEQUENCE_BATCH_GRAPH", "PARTIAL_GRAPH_KEY", "RUNTIME_PACKAGE_INSTALL",
     "PRECISION_DOWNGRADE", "SCORE_PHYSICS_FLOOR", "SCORE_IMPOSSIBLE", "SCORE_SUSPECT_FLOOR",

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "kernelguard"
-version = "0.3.0"
+version = "0.3.1"
 description = "Rule-based GPU kernel hack detector."
 readme = "README.md"
 requires-python = ">=3.11"