Promote shape-aware r14 EXL3 prefill

[malaiwah]

What changed

• Compose vLLM #210's bounded/sliced EXL3 prefill arena before vLLM #219's
  shape-aware mixed-K executor in the fail-closed field-review manifest.
• Keep the cooperative one-grid/block-8 route for decode-sized rows and use
  serial homogeneous K3/K4 block-64 plans for large prefills.
• Select a 1,024-row reusable arena for the 3.25-bpw profile and document its
  measured memory/performance trade.
• Update the cache ABI namespace, patch checksums, profile tests, README,
  changelog, self-service guidance, and full AIBeast qualification record.

Why

GG v20-r14's one-grid mixed-K route regressed 3K/32K/128K prefill by about 23%
versus r13. The executor fix recovers 28.4-29.3% over pristine r14 while
retaining r14's hardening and decode route. The 1,024-row #210 arena then
returns about 665 MiB/GPU. A 1,536-row arm returned only about 332 MiB/GPU but
paid the same roughly 10-11% PP cost, so 1,024 is the useful safety/performance
point for the full 524,288-token profile.

Live qualification

Source-exact turnkey image on AIBeast: 4x RTX PRO 6000 Blackwell 96 GB at
280 W, driver 595.71.05, CUDA 13.2; willfalco 3.25-bpw snapshot
61d2b6b757f6a4ac7098a78d861f2033497532dc; TP4/DCP4; 2,048 GPU blocks;
125 GiB DRAM plus bounded 512 GiB NVMe LMCache.

• cold unique-prefix PP: 2,121 / 1,932 / 1,837 tok/s at 3K / 32K / 128K
• aggregate MTP5 TG: 93 / 126 / 162 / 240 tok/s at C1 / C2 / C4 / C8
• MTP5 MAL: 3.31 / 3.17 / 3.11 / 3.70
• actual 509,022-token prompt: 3/3 needles, no degeneration
• 524,288 active GPU-KV tokens; no request failure, preemption, exception, or
  CUDA OOM
• both GLM-5.2 and local-primary aliases passed

The turnkey profile itself retains MTP3 because the matched field-review
workload found better acceptance and tail latency at depth 3; MTP5 was used in
the AIBeast control so only the executor/capacity stack changed.

Upstream lineage

• capacity/slicing: [GG] Bound EXL3 Trellis prefill arena capacity local-inference-lab/vllm#210
• shape-aware executor: exl3: use shape-aware mixed decode and serial prefill local-inference-lab/vllm#219
• future mixed block-64 kernel: EXL3: one-grid mixed K3/K4 regresses large-M SM120 prefill local-inference-lab/b12x#107
• downstream rotation/scratch prototype (distinct):
  fix(exl3): share target and draft prefill arena #15

Validation

• exact repository CI suite from AGENTS.md
• all three configuration-smoke profiles
• field-review manifest/application tests
• EXL3 mixed-K and parity-ABI patch tests
• structured-output patch and field-review log-audit tests
• JSON and whitespace checks
• source-exact GPU image boot, cold performance matrix, alias checks, LMCache
  activity, and near-maximum retrieval gate

[malaiwah]

Post-merge production handoff on AIBeast is green:

• glm52-turnkey-r14-cap1024-shape-prod is healthy on port 8000; port 18000
  is closed and no qualification container owns GPU memory.
• Warm compile-cache reuse took 0.60 s for the backbone; graph capture took
  8 s and 0.20 GiB/GPU. The runtime reported the qualified 1,024-row target
  and draft plans and exactly 524,288 GPU-KV tokens.
• The built-in authenticated verifier passed short prompts, strict JSON with
  thinking, and 3/3 needles at 32,853 tokens. A separate local-primary
  verifier passed arithmetic, factual, instruction, and structured-output
  gates.
• Under immediate agent traffic, generation windows reached 120-125 tok/s at
  C1 with MTP5 MAL 4.27-4.45; external prefix-cache hit rate was 18.6%.
• Full boot and soak logs contain no ERROR, traceback, CUDA OOM, request
  failure, or preemption.
• Checkpoint remains read-only; LMCache remains 125 GiB DRAM + bounded 512 GiB
  local NVMe.

The host's Podman release cannot add a restart policy to an existing container
in place, so I deliberately did not replace the healthy service solely for
that non-runtime setting. PID 1 still supervises/restarts the vLLM process;
after a host reboot the standard launch command is required.