Phase-2: split cap (constants only) + users>=2 filter (activations)

Two refinements after the L2 reservation fix:

1. The 2 KB cap (Phase-1 \xc2\xa710.1) now applies ONLY to ConstantBuffers.
   Activations are dynamically allocated in MEMORYARENA_L2 (not
   linker-placed static PI_L2), so the per-tile static-read codegen
   bug \xc2\xa710.1 was protecting against does not apply to them.
   Removing the cap on activations lets large multi-consumer
   activations (residual splits, attention K/V) actually promote.

2. Activations now require users >= 2 (multi-consumer) by default.
   Single-use activations gain nothing structural from promotion
   (the next-layer cl_ram round-trip happens once either way) and
   triggered an L2-allocator corner case on AnomalyDetection
   (35/640 errors when 9 users=1 onnxGemm_* tensors entered the
   reservation pool).  Constants bypass this filter.

Phase-2 sweep with both refinements (and reservation fix from a3fc5d1):

| Model           | const-only (P1) | act-included (P2) | Δ vs baseline |
|-----------------|----------------:|------------------:|--------------:|
| IC100           |   2 389 072 cyc |   2 277 046 cyc   | -8.4 %%       |
| ml1             |   3 475 840 cyc |   3 095 426 cyc   | -26.8 %%      |
| miniMobileNet   |     134 946 cyc |     134 946 cyc   | -29.6 %%      |
| AnomalyDet 200  |     509 558 cyc |     509 464 cyc   | -15.3 %%      |

All 4 PASS output equality.  The users>=2 filter costs perf vs the
no-filter run on ml1 / miniMobileNet (single-use activations were
genuine wins there) but restores AD correctness.  Future Phase-2
work: figure out why AD's single-use activations specifically trip
the reservation allocator and lift the filter only for safe models.

Author: runwangdl

Deeploy/MemoryLevelExtension/OptimizationPasses/MemoryLevelAnnotationPasses.py

@@ -234,9 +234,28 @@ 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
-            if self.maxBufferBytes is not None and _bufferSizeBytes(buf) > self.maxBufferBytes:
-                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
             reuse = max(1, len(getattr(buf, "_users", [])))
+            # For activations specifically (non-Constant VariableBuffer),
+            # require multi-consumer reuse (users >= 2).  A single-use
+            # activation gains nothing from promotion (the next-layer
+            # cl_ram round-trip happens once either way) and adds reservation
+            # pressure that triggered AnomalyDetection corner cases at the
+            # L2 allocator (Phase-2 §7.5 follow-up).  Constants bypass this
+            # filter — small reuse=1 weights still benefit because they're
+            # loaded once at network init, not per-inference.
+            if not isinstance(buf, ConstantBuffer) and reuse < 2:
+                continue
             if reuse < self.minReuse:
                 continue
             size = _bufferSizeBytes(buf)