[ashhart]

This carries @gilby's #1335 forward at their request: their commit with their authorship, rebased onto main, with the opt-in they asked us to add. Part of #1334.

• Off by default: TF_QWEN_REQUEST_STARTS=1 turns on exact message cuts for eligible Qwen families. Unset, zero, empty or anything else behaves as main does, with a test.
• Beside shared KV it's refused by name. Exact cuts would need partial-page snapshots, and that's a step of its own.

Host tests pass and lean stays at 425. Not yet run on a GPU: the six-turn tables on a GB10 against main are ours, and it won't land until those are posted.

[ashhart]

/atc exchange 18b05daeca073604fd3bf7e24e29a1372246863a5edc5b4d8f2c29ce5e617eb2

[ashhart]

This now takes the per-family switch from @laidick's #2404 (commit 90e7309ba): requestStep reads the switch the family names (request_starts_switch), so turning on Qwen's TF_QWEN_REQUEST_STARTS sends message starts only to Qwen's two CUDA entry points, a family that names no switch never gets them, and with every switch unset every CUDA family is on main's path. Qwen's switch name, its refusal with TF_QWEN_KV_SHARE and its tests are unchanged; a host test walks all twelve registered CUDA families. When #2404 rebases onto this, its copy of that plumbing can go. This PR still waits on its six-turn tables.

[ashhart]

The six-turn tables are in: one GB10, Qwen3.8-27B MLX 4-bit with the DFlash2 drafter, main's merge base 09e2f7fa6 against this head 90e7309ba. Each arm ran two fixed six-turn chats, a short one (1.5K tokens) and a long one (21.4K tokens), on a warm server, and then each turn again on a cold server.

Replies are equal on every arm and turn. Unset, the switch runs main's path: new-off is level with base everywhere. TF_QWEN_REQUEST_STARTS=1 beside TF_QWEN_KV_SHARE above 0 is refused by name before serving.

First-token time on the warm server, turns 2 to 6:

Chat	main	this PR, unset	this PR, TF_QWEN_REQUEST_STARTS=1
Long (21.4K)	0.93-1.06 s, 20,480 cached	0.93-1.07 s, 20,480 cached	0.17-0.27 s, 21,371 cached
Short (1.5K)	1.16-1.36 s, 0 cached	1.16-1.35 s, 0 cached	0.11-0.22 s, 1,450 cached

Main resumes on the 2,048-row grid, so the short chat never resumes and the long one re-prefills about 900 rows a turn. With the switch on, both resume within a few rows of turn 1's end. Turn 1 costs 0.04 s more on the long chat and 0.17 s on the short one, for the extra cuts. Cold servers are level with main in every arm (16.2-16.4 s and 1.1-1.4 s).

One thing the tables show: cached_tokens stays at turn 1's history end (21,371 and 1,450) for turns 2 to 6. Later turns resume from turn 1's state and prefill the rest, so the remaining prefill grows by about 50 rows a turn rather than staying a few rows. It's still exact and still well ahead of main, but it isn't "the next turn resumes where its history ended" for turn 3 onwards. We'll look at whether a later turn's own cuts should be kept too, as a follow-up. The opt-in itself is ready to land once it's rebased onto main.