Phase-2: apply 2 KB cap to BOTH constants and activations - CCT_2 now PASSES

Earlier 'cap-on-constants-only' (a8694f7) made CCT_2 fail 10/10
because its 7x 32 KB residual / LayerNorm-output activations hit the
write-side mirror of the §10.1 codegen bug (writer layer tile-loops
output to a fixed L2 base address, every tile clobbers the previous).

The §10.1 bug is symmetric: any buffer >2 KB tiled across the
producer's output dimension corrupts under promotion, regardless of
whether it's a static-PI_L2 ConstantBuffer (read-side) or a
dynamically-allocated MEMORYARENA_L2 VariableBuffer (write-side).
Apply the cap uniformly.

Cost: IC100 / miniMobileNet lose their large-activation gains
(x3_tensor_split 16 KB blocked), back to const-only level.
microLlama1 retains its activation gains (all activations <2 KB
already).  All 5 models PASS:

| Model           | const-only      | act+cap+users>=2  | Δ vs baseline |
|-----------------|----------------:|------------------:|--------------:|
| IC100           |   2 389 072 cyc |   2 389 072 cyc   |  -3.9 %%      |
| microLlama1     |   3 475 840 cyc |   3 095 426 cyc   | -26.8 %%      |
| miniMobileNet   |     134 946 cyc |     134 946 cyc   | -29.6 %%      |
| AnomalyDet 200  |     509 558 cyc |     509 738 cyc   | -15.3 %%      |
| CCT_2           | 309 888 350 cyc | 309 893 862 cyc   | ~0 %% (PASS)  |

Correctness wins over peak perf for IC/mmn — recovering the
large-activation gains there requires the proper Phase-3 fix
(per-tile L1↔L2 source/destination advancement when the buffer is
at the closure's externalMemory level).

Author: runwangdl

Deeploy/MemoryLevelExtension/OptimizationPasses/MemoryLevelAnnotationPasses.py

@@ -234,17 +234,22 @@ def apply(self, ctxt: NetworkContext, graph: gs.Graph) -> Tuple[NetworkContext,
                 continue
             if self.excludeNamePatterns and any(pat in name for pat in self.excludeNamePatterns):
                 continue
-            # Per-buffer-size cap: applies to ConstantBuffer ONLY.
-            # ConstantBuffers above the cap are subject to the §10.1 read-side
-            # codegen bug (per-tile static PI_L2 read with no source pointer
-            # advancement); the cap is the safe interim workaround.
-            # Activations live in MEMORYARENA_L2 (dynamically allocated, not
-            # linker-placed static PI_L2), so the §10.1 mechanism does not
-            # apply to them — large activations are free to promote provided
-            # the users >= 2 filter below is satisfied.
-            if isinstance(buf, ConstantBuffer):
-                if self.maxBufferBytes is not None and _bufferSizeBytes(buf) > self.maxBufferBytes:
-                    continue
+            # Per-buffer-size cap.  Applied to BOTH ConstantBuffer (because
+            # of the §10.1 read-side codegen bug for static-PI_L2 weights
+            # tiled across an output dimension) AND VariableBuffer (because
+            # of the symmetric write-side codegen bug for promoted
+            # activations whose writer tiles across them — observed on
+            # CCT_2's 32 KB residual / LayerNorm-output activations).
+            # Both bugs leave the per-tile L1↔L2 source/destination pointer
+            # at a fixed offset across the tile loop, so any buffer larger
+            # than a single L1 tile slab corrupts under promotion.  The
+            # 2 KB default is the empirically-anchored safe boundary for
+            # this codebase (see Phase-1 §10.1 bisection).  Activations
+            # smaller than the cap (residual splits, attention K/V tiles
+            # that fit in one L1 tile, small per-channel intermediates)
+            # still promote and yield real cycles.
+            if self.maxBufferBytes is not None and _bufferSizeBytes(buf) > self.maxBufferBytes:
+                continue
             reuse = max(1, len(getattr(buf, "_users", [])))
             # For activations specifically (non-Constant VariableBuffer),
             # require multi-consumer reuse (users >= 2).  A single-use