test(cuda): CudaGraphCaptureEquivalenceTest.EagerVsGraphDecode_QuantizedKv_Match too-tight tolerance causes full-suite-only failures #174

Context

Documented pre-existing flake (memory: known to pass standalone, fail in the full CUDA suite,
confirmed 5x across sessions) — surfaced again this session while shipping #170. Investigated per
user request to "explore the isolation flake."

Root cause (confirmed, not a real bug)

EagerVsGraphDecode_QuantizedKv_Match asserts maxDiff < 1e-3f on first-step logits, with a
comment claiming eager and graph paths are bit-identical since they invoke the same kernels.

Its sibling test right above it, EagerVsGraphDecode_Match, already documents and tolerates
(< 5.0f) the real cause: PTX-JIT cache state left behind by other CUDA tests in the same process
(module/kernel registration order shifts SASS scheduling, which shifts FP accumulation order in
otherwise-identical arithmetic) perturbs logits by up to ~0.25 max-abs in the full suite, with
argmax unaffected. The QuantizedKv test's stronger "bit-identical" claim just isn't true once other
tests have run first — reproduced directly this session:

• Standalone: Step 0 logit max abs diff: 0.000000
• Full DotLLM.Tests.Unit.Cuda suite: Step 0 logit max abs diff: 0.201172 — well inside the
  sibling test's own accepted tolerance, comfortably failing the QuantizedKv test's 1e-3f.

Fix

Loosen the tolerance to 5.0f (matching the sibling test exactly) and correct the comment to
describe the real, benign cause instead of the disproven "bit-identical" claim.

Verification

Full DotLLM.Tests.Unit.Cuda suite, same conditions that previously reproduced the failure:
before fix — 1 failed (this test); after fix — 339 total, 303 passed, 36 skipped, 0 failed.