Reorganise: group 313 tasks into 17 families under tasks/, generators under tools/

tasks/distributed-multi-gpu/dist-allgather-gemm/instruction.md

@@ -71,6 +71,17 @@ cat /logs/verifier/details.json      # correct / speedup / tflops
 
 Get `ag_gemm` **correct** first (match the torch reference), then optimize the overlap for TFLOP/s.
 
+## Know your hardware — do not assume it
+
+The exact GPU is deliberately not stated. Query it at run time with `torch.cuda.get_device_properties(0)`:
+`.multi_processor_count` (size any persistent grid from this, never a constant),
+`.shared_memory_per_block_optin`, register and thread limits, and the compute capability. `nvidia-smi --query-gpu=name,memory.total,clocks.max.sm --format=csv`
+gives clocks, and `ncu` reports achieved occupancy and bandwidth directly.
+
+Measure achieved HBM bandwidth with a large stream-copy rather than trusting a datasheet figure. Do not
+hardcode an SM count, a tile size derived from one, or a peak FLOP/bandwidth constant — the same
+submission is graded on whatever device it lands on.
+
 ## Constraints
 
 - **Offline.** pypi / github / huggingface are firewalled — you cannot fetch the original kernel. Build the

tasks/distributed-multi-gpu/dist-tp-fused-kernels/instruction.md

@@ -62,6 +62,17 @@ cat /logs/verifier/details.json      # per-kernel TFLOP/s + correctness
 Get one kernel correct end-to-end first, then optimize its overlap, then bring up the others. Correctness on
 all three is required before you score anything.
 
+## Know your hardware — do not assume it
+
+The exact GPU is deliberately not stated. Query it at run time with `torch.cuda.get_device_properties(0)`:
+`.multi_processor_count` (size any persistent grid from this, never a constant),
+`.shared_memory_per_block_optin`, register and thread limits, and the compute capability. `nvidia-smi --query-gpu=name,memory.total,clocks.max.sm --format=csv`
+gives clocks, and `ncu` reports achieved occupancy and bandwidth directly.
+
+Measure achieved HBM bandwidth with a large stream-copy rather than trusting a datasheet figure. Do not
+hardcode an SM count, a tile size derived from one, or a peak FLOP/bandwidth constant — the same
+submission is graded on whatever device it lands on.
+
 ## Constraints
 
 - **Offline.** pypi / github / huggingface are firewalled. torch + `triton_dist` + NVSHMEM are pre-installed.