feat(deepseek-v4): enable bf16 RL, bump vLLM to 0.29.0 and prime-envs

[garrett361]

DeepSeek V4 Flash passes the KL mismatch merge bar for bf16 RL. The blocker was a vLLM weight-reload bug that left captured CUDA graphs reading a freed pointer, fixed upstream in vllm#52626 and picked up here by pinning vLLM 0.29.0.

math environment, batch_size = 64, 20 steps, lr = 0, full depth, 4 trainer nodes + 1 inference node:

step	replay on	replay off
1	0.00037	0.00091
2	0.00030	0.00097
3	0.00040	0.00108
4	0.00025	0.00092
5	0.00028	0.00068
6	0.00031	0.00083
7	0.00037	0.00097
8	0.00030	0.00070
9	0.00040	0.00077
10	0.00030	0.00101
11	0.00031	0.00067
12	0.00035	0.00099
13	0.00028	0.00085
14	0.00034	0.00075
15	0.00035	0.00065
16	0.00030	0.00108
17	0.00037	0.00094
18	0.00037	0.00081
19	0.00029	0.00075
20	0.00039	0.00110

All 40 entries under the 0.015 bar: max 0.00040 with replay on, 0.00110 with replay off.

reverse-text

The same model and topology on reverse-text at batch_size = 32, over 20 steps:

configuration	mismatch_kl	max
vLLM 0.28, router replay off	0.194	0.490
this PR, router replay off	0.024	0.053
this PR, as shipped	0.0016	0.0027

The vLLM 0.28 row is a 2026-09-14 measurement carried forward. The two this PR rows were
re-measured on the current branch, so rerunning the baseline would compare against a different
main than the one it was originally a baseline for.

• The math table was measured before this branch bumped prime-envs. At the old pin 1f1e050ab, environments/math/ set network_allow = [], which makes verifiers refuse the subprocess runtime every config in this repo uses, so the run needed a scratch wrapper taskset restoring the wildcard. The bump to a6fc0c66f picks up prime-envs Simple implementation of different loss scaling #795, which reverts that default-deny, so the environment now runs unwrapped. The other two measurement-only settings, constant_trainer_batch_size = false and an inert length-penalty tiebreaker so frozen-policy groups ship, are still not committed.
• Serving the published repo would make vLLM allocate fp8 linears and MXFP4 experts that bf16 weights cannot fill, since its config.json carries quantization_config and expert_dtype = "fp4". These configs use PrimeIntellect/DeepSeek-V4-Flash-0731-bf16 instead, currently private.