[MUH] Fix 7 structural discrepancies vs CCCL — read source, not grep

Fixes found by reading all 17 muh files + 6 CCCL counterpart
policy_selectors as full source code input:

1. topk: BLOCK_LOAD_DIRECT → BLOCK_LOAD_VECTORIZE (CCCL SM90+ uses
   VECTORIZE). bits_per_pass was wrong (muh: ks<=4→9, CCCL: ks>=2→11).
   items now computed dynamically (4*4/key_size) not hardcoded.

2. reduce: added determinism dispatch — three modes matching CCCL:
   gpu_to_gpu (BLOCK_REDUCE_RAKING, vec_size=1, LOAD_DEFAULT),
   run_to_run (WARP_REDUCTIONS, LOAD_LDG, default),
   not_guaranteed (WARP_REDUCTIONS_NONDETERMINISTIC).
   Added bi100_det_float32 and bi100_det_float64 tuning structs
   with SM90 benchmark reference values.

3. batch_memcpy: flat single-tier → SmallBuffer+LargeBuffer two-tier
   matching CCCL structure (128 threads small, 256 threads large,
   warp_threshold=128, block_threshold=8192).

4. transform: single BulkPolicy → three-policy structure
   (VectorizedPolicy + AsyncCopyPolicy + PrefetchPolicy) matching CCCL.
   items_per_thread computed from bytes_in_flight / (threads * elem_size).

5. compile_test: 17 checks → 33 checks. Now verifies exact values:
   reduce determinism modes, topk VECTORIZE + bits=11, batch_memcpy
   two-tier thresholds, transform three-policy structure.

6. gen_patch: added fallback extraction for inline policy_selector
   values (topk now generates SAMPLING_BLOCK_SIZE patch).

7. MUH_PROJECT_CHECKPOINT.md: 'PRD设计阶段还没有代码' → actual status.

7 files changed, 413 insertions, 265 deletions.

muh/gen_patch.py

@@ -116,6 +116,36 @@ VLLM_INJECTION_POINTS = {
 }
 
 
+def extract_hardcoded_values(filepath):
+    """Fallback: extract key values from policy_selector return statements.
+    
+    For algorithms where bi100_* structs don't exist (values computed inline).
+    Extracts the threads_per_block from the first return in the iluvatar branch.
+    """
+    with open(filepath, 'r') as f:
+        content = f.read()
+    
+    algo = algo_from_filename(filepath)
+    
+    # Find iluvatar branch
+    iluvatar_match = re.search(
+        r'hw\.at_least\(.*iluvatar.*?\)\s*\{(.*?)(?=\n\s{2,4}\})',
+        content, re.DOTALL
+    )
+    if not iluvatar_match:
+        return []
+    
+    branch = iluvatar_match.group(1)
+    
+    # Find return {N, ...} — first integer is typically threads_per_block
+    return_match = re.search(r'return\s*\{(\d+)', branch)
+    if not return_match:
+        return []
+    
+    threads = int(return_match.group(1))
+    return [('__inline__', {'threads': threads})]
+
+
 def generate_patches(header_dir):
     """Read all tuning headers, extract bi100 values, generate patches."""
     patches = []
@@ -131,8 +161,11 @@ def generate_patches(header_dir):
         structs = extract_bi100_structs(hpath)
 
         if not structs:
-            summary.append(f"SKIP {algo}: no bi100_* structs found")
-            continue
+            # Fallback: try extracting inline values from policy_selector
+            structs = extract_hardcoded_values(hpath)
+            if not structs:
+                summary.append(f"SKIP {algo}: no bi100_* structs and no inline values found")
+                continue
 
         # Use the first non-default struct as the primary tuning
         # (default is fallback; prefer the type-specific ones)