[xvcezrqpbxtpzblrt]

Problem

A 512K context on a 16 GB card could not fit in one run. The prompt path borrows part of the expert cache as a staging buffer (the loan), and the K/V's ceiling stopped at the loan's first slot: the borrowed region was excluded from the cache for the whole run. When the K/V needed those slots, the engine asked the driver for new VRAM, and on a card that has none left the request ended with and new VRAM, the cache being at its floor.

Result (measured)

k8v4, 16 GB RTX 3080 Ti Laptop, --max-context 524288, 510K prompt, 1.5 GiB reserve, --prefill 6144, staged chunks:

run	prompt	prefill	K/V	new VRAM
1	522,000	846.8 tok/s	5.30 GiB	none (log checked)
2	522,000	831.7 tok/s	5.30 GiB	none (K/V grown line has no and new VRAM)

The loan shrinks from 6,144 to 4,608 tokens and the freed slots become K/V cells; nothing is requested from the driver. 128K on a cool card: 1463.5 tok/s (previous build 1514.5, within run-to-run heat).

What changed

• kvg_ensure, at its floor and before any new VRAM: asks the prompt path to re-lay out in the largest chunk whose freed granules cover the K/V's need, then unmaps the tail into the cache (kvg_fit_loan).
• New option --prefill-stream-min N: chunks of N tokens or more stream every expert instead of staging the routed ones; an explicit value wins over the CPU share's own limit. Default behaviour is unchanged (the built-in 1,024 threshold).
• Short contexts never take the new path (remain is 0), so their behaviour is the same as before.

Validation

• 510K runs above, full logs kept (K/V grown without and new VRAM).
• 32K: 1,672.7 tok/s; 128K: 1,463.5 tok/s (cool card).
• Not run: 8 GB and 24 GB cards, multi-GPU (the path is single-participant), HIP, SYCL (this source has no such path), Metal (elastic K/V is off there).

Limits

• CUDA and HIP only; the SYCL tree is not touched by this change.
• Requires --kv-grow (the elastic K/V path).
• On a 1.5 GiB reserve, int8 at 510K still asks for about 1.04 GiB of new VRAM; k8v4 does not. int8 needs a 1 GiB reserve to stay inside the cache.

[xvcezrqpbxtpzblrt]

Superseded by #4606, which carries this change with the rest of the 0.2.7 branch. Closing so the review has one place.