kv: replace eligible dense causal masks with compact prefixes

[GenerelSchwerz]

Status

Official isolated head: ea066a68ad6eb6db701251d5ae23dfea3948ee26
Measured runtime head: 222eec2e7c91296b5f9335389dcff0cfc8a08129
Exact base: 4b86269fdf001de44dd96e9c9ae26a9e25091cca (beellama/main)

This PR is ready for review after a normal merge of current beellama/main.
It remains unmerged. The evidence-only commit after the runtime head changes no
binary source.

Scope

Eligible single-stream contiguous standard-KV graphs pass the existing I64
write-index tensor directly to FlashAttention as an exclusive causal-prefix
descriptor. Unsupported layouts fail closed to the existing dense mask.

• no public flag, user CMake option, model/architecture special case, context
  threshold, or prompt/decode policy;
• no native-Q8, KVarN implementation, workspace, allocator-policy, VMM-policy,
  host-staging, speculation, or development-branch feature;
• CPU and CUDA descriptor consumers, with CUDA capability registration;
• compile-time <..., false> dense and <..., true> compact vector
  specializations retained across the existing cache-format matrix;
• explicit block-owned shared-memory lifetime for the compact MMA first bound;
  and
• one optional graph value owns representation presence and retained K span,
  removing an invalid boolean/count state without hot-path work.

The main merge is c542f137c; the isolated lifetime repair is 1a3e6fcf3.
Focused compiled-capability coverage replaces compact-specific source-text
assertions. The retained two-turn runner now takes model and artifact paths as
arguments rather than embedding a machine model path.

Correctness and quality

• Fresh Release CUDA/FA SM120a builds, default quant matrix, KVarN off, maximum
  12 build jobs under /tmp/beellama-cuda-build.lock.
• Compiled CUDA dense-versus-compact oracle: 6/6, including D=256 F16 MMA
  with no GQA and GQA 6.
• Focused graph, allocator, vector-generation, route, fallback, parser, and
  replay CTests: 10/10.
• Main/candidate/main PPL with -b 512 -ub 256 --chunks 4: identical chunks
  1.9315, 2.1279, 2.2498, 2.1674 and final
  2.1674 +/- 0.03849; measured increase zero.

Every GPU lifecycle used the complete
flock /tmp/beellama-single-gpu.lock -c 'COMMAND' form, fresh processes, and
native progress. The one 128K placement setup failure is explicitly invalid
and contributes no result.

Matched runtime and resources

Qwen3.8 27B, q8_0 K/V, batch/ubatch 512/256, RTX 5070 Ti, CUDA 13.3, driver
610.57.04:

gate	compact effect
4K GPU-KV prefill	-0.0736% time-adjusted; neutral
4K GPU-KV decode	+0.0482%
30K GPU-KV prefill / decode	+0.7765% / +0.4059%
128K CPU-pinned-KV prefill / decode	+1.2745% / +0.4923%

The user-requested closing A4 anchor makes 4K prefill chronological
A B A B A B A. Baseline falls monotonically from 1810.6711 to 1798.3230
tok/s. A linear elapsed-time/source model estimates -1.3282 tok/s (-0.0736%)
with 1.7246 tok/s standard error; local interpolated effects are -0.1665%,
-0.0330%, and -0.2066%. This is neutral within observed time/process variation,
not a speedup and not a detected regression. The 128K decode series retains all
four processes/source; its +0.4923% aggregate has wide process variance and is
not promoted to a strict speedup claim.

Independent 100 ms resource sampling measured process-VRAM deltas of
-2 / 0 / -58 MiB at 4K/30K/128K prefill. At 128K, total compute backing
falls 123.5 MiB, while pinned KV context, resident KV, ordinary-host buffers,
and CUDA VMM high-water are unchanged. Allocation-class telemetry and the
process sampler are separate from timing.

Evidence

The durable design, commands, identities, limitations, and results are in
docs/compact-causal-mask-vram-investigation.md. The 135-file local artifact
manifest has SHA-256
d4246230238f39cf643a2a2a1037835cc4a20dd490844abc1dc976d1180206a1.

The final main-relative audit is 27 files and contains no PR 4, workspace,
allocator-policy, VMM-policy, server, or development-only source. This PR must
not be merged by the readiness maintainer; normal review owns landing.