[ashhart]

Changelog only. Three lines under Unreleased that tonight's landings left short:

• cuda: stage synchronous downloads through a thread's pinned buffer #2103 landed without a line. It gets one: on integrated GPUs such as the GB10, synchronous downloads stage through a pinned buffer, with the cost from the PR's own table, 212 µs instead of 85 µs at 4 MiB on an idle GB10.
• gemma4 CUDA: Gemma 4 E4B (NVIDIA Nemotron 3.5 Bolt E4B), per-layer inputs and shared KV #2230's line gets its thanks to @jschmied.
• Rerank on HIP: compute the shared trunk once a pass (25% faster, same bytes) #2262's line cites its own PR, not just Qwen3 rerank on HIP: speed levers from the R9700 profile (tracker) #2245, and uses the PR's token count: 16 documents of 323 tokens, 87 of them shared, not "about 260".