Optimize GLM 5.3 prefill and add backend-neutral metrics

[nessie993]

Closes the current prefill/tok/s tuning pass from #263 while keeping GLM B outside Ramjet routing and Qwen A unchanged.

What changed

• promote the measured GLM SM120 defaults to chunked-prefill=6144, max-prefill-tokens=6144, and max-total-tokens=500000
• add a guarded B-only transition owner with deployment locking, 600W inference-ceiling admission, Qwen A identity checks, deferred inference budget, and immediate dead-engine detection
• make context_frontier.py reconcile backend-neutral request, prompt, cached-prompt, and generation counters for SGLang and vLLM
• add forced-length decode support so per-stream decode samples cannot end early at EOS
• record the power A/B, 8K rejection, clean 6K frontier/c4 results, and current upstream engine research

Live result on node06

8K was rejected twice during built-in serving warmup: sparse attention needed a 256MiB output with 124.94MiB free at the full 524,288-token pool and 244.94MiB free at 500,000. Both failures occurred before readiness or benchmark traffic and stopped only GLM B.

6K/500K became ready in 1,036s and passed its smoke. Initial text shapes that produced late Triton loads were treated only as warmup. Subsequent clean, guarded, exactly reconciled cells measured:

cell	4K	6K	change
2K cold prefill	5,174 tok/s	5,174 tok/s	0.0%
8K cold prefill	5,415 tok/s	5,817 tok/s	+7.4%
32K cold prefill	5,897 tok/s	5,961 tok/s	+1.1%
64K cold prefill	5,884 tok/s	5,951 tok/s	+1.1%
c4 aggregate output	370.7 tok/s	388.2 tok/s	+4.7%
c4 per-stream decode	106.3 tok/s	109.4 tok/s	+2.9%

The accepted c4 cell completed 20/20 requests and reconciled all 5,120 generated tokens. Intake was 42C; peak GLM GPU temperature was 62C in c4 and 73C in the long-context frontier. The accepted setting gives up 24,288 shared token slots (4.6%), so the recipe and journal state that trade explicitly.

The promoted Compose was staged without recreating the already-matching live engine. GLM B remains loopback/LB-isolated on GPUs 4-5; Qwen A remains healthy and unchanged; GPUs 6-7 remain free.

Engine research

• vLLM v0.29.0 still cannot run this SM120 NoPE quant on a reviewed immutable path: [Bugfix][SM120][MLA] Support NoPE sparse MLA (GLM-5.3-Flash) on the FlashInfer SM120 backend vllm-project/vllm#55277 remains open/dirty
• vLLM FlashKDA and masked-MHA prefill work is promising but not deployable yet: [Perf][GLM-5.3-Flash] Use FlashKDA for KDA chunked prefill (1.7-3.8x faster than the Triton chunk path) vllm-project/vllm#55737 and [Perf][GLM-5.3-Flash] Dense/masked-MHA sparse prefill for the NoPE (256, 0, 256) layout + skip the NoPE K concat vllm-project/vllm#55738
• SGLang small-batch SM120 BA projection is still blocked: [Perf] Add opt-in SM120 BF16 BA projection for small batches sgl-project/sglang#38635
• SGLang K-pool metadata fusion is already enabled in this canary and has no isolated upstream speed claim: Reduce repeated attention setup during speculative decoding sgl-project/sglang#38213
• compact NoPE storage is a capacity candidate, not a measured speedup, and remains open: Enable compact GLM NoPE KV storage with native FlashInfer sgl-project/sglang#38430

Validation

• cargo fmt --check
• cargo clippy --locked --all-targets --all-features -- -D warnings
• cargo test --locked
• cargo build --release --locked
• python3 bench/agentbench.py validate
• PYTHONPATH=bench python3 -m unittest discover -s bench -p 'test_*.py' (723 tests)
• python3 deploy/glm53_flash_sm120/validate-compose.py
• bash -n deploy/glm53_flash_sm120/node06-prefill-6144-rollout.sh
• git diff --check