Charge a copy held in its source's memory only what it adds, in the model and its capacity screen

[asyms]

The memory model counted a copy only beyond its source when the whole transfer stayed in one memory. A multicast that also reaches other memories paid in full for the copy it lands next to its source, though that reader reads it in place: in the attention head on the TPU-like quad core, the exp output and its copy for the div were both counted on the 128 KB vector core, which made the head infeasible. The capacity screen also counted such copies in full while the model it guards did not.

_memory_loads now charges any copy held in its source's memory only what it holds beyond the source, and capacity_screen applies the same rule to pinned tensors. A transfer that stays in one memory is charged as before.

Test: the attention head solves on the TPU-like quad core. It fails without the change.

Stacked on #156.

[github-actions]

Stream AIE Metrics Regression Guard

⚠ 5 cell(s) flagged (total_latency > 0.1% tol): hardware_two_conv[eyeriss_like_dual_core], hardware_two_conv[eyeriss_like_quad_core], hardware_two_conv[meta_prototype], hardware_two_conv[simba_small], hardware_two_conv[tpu_like_quad_core]

16 of 16 cells captured

Provenance: baseline 32b8eaa94931143f86c3a135c774febd1d8b96ae | date 2026-10-04 | Python 3.12.3 | backend ortools_gscip
Note: mip_gap: null — OR-Tools GSCIP
Columns: array fill = per-layer PE-array spatial fill (dataflow quality); MAC eff (e2e) = useful MACs / (chip peak MACs/cycle × total latency), the true fraction of the chip's compute used (incl. idle cores, temporal stalls & transfers).

hardware_swiglu — 8 hardware

seq_len=256, embedding_dim=2048, hidden_dim=8192, bf16; layer-fused tiles seq=16/embedding=128/hidden=32

Hardware	total_latency (base → cur)	Δ%	array fill	MAC eff (e2e)	note
eyeriss_like_dual_core	149684492 → 149684492	+0.00%	52%	26%
eyeriss_like_quad_core	101581557 → 101581557	+0.00%	52%	19%
eyeriss_like_single_core	303038604 → 303038604	+0.00%	52%	25%
fusemax	233766951 → 233766951	+0.00%	9.3%	0.26%
meta_prototype	101711933 → 101711933	+0.00%	100%	6.2%
simba	125344507.60 → 125344507.60	+0.00%	65%	1.0%
simba_small	106627594 → 106627594	+0.00%	65%	3.0%
tpu_like_quad_core	101580877 → 101580877	+0.00%	97%	3.1%

hardware_two_conv — 8 hardware (⚠ 5 flagged)

batch=1, in_ch=8, H=32, W=32, out_ch1=16, out_ch2=32, kernel=3x3, bf16 (generic auto-tiling, not layer-fused)

Hardware	total_latency (base → cur)	Δ%	array fill	MAC eff (e2e)	note
eyeriss_like_dual_core	74631 → 74761	⚠ ↑+0.17%	51%	23%
eyeriss_like_quad_core	39917 → 40209	⚠ ↑+0.73%	51%	22%
eyeriss_like_single_core	115003 → 115003	+0.00%	68%	31%
fusemax	187612 → 187676	↑+0.03%	1.4%	0.05%	low array utilization
meta_prototype	15341 → 15487	⚠ ↑+0.95%	51%	19%
simba	4610.67 → 4610.67	+0.00%	50%	4.9%
simba_small	8706 → 8998	⚠ ↑+3.35%	50%	16%
tpu_like_quad_core	12300 → 12592	⚠ ↑+2.37%	42%	11%

To regenerate baseline: python scripts/analysis/render_metrics_comment.py --update-baseline