fix(MODEL-TEXT-qwen3-qwen3-for-causal-lm): diagnose the CPU paged anchor drift

Row: MODEL-TEXT-qwen3-qwen3-for-causal-lm

The CPU test_qwen3_paged_engine gate fails on unchanged source at 760c8dd7e27781430b909c3679790c52fe332182. Both the ordinary Qwen3-0.6B case and its KV-boundary case fail the exact local anchor check at tests/parity/test_qwen3_paged_engine.cpp:365.

Prompt 0 is The capital of France is. Greedy generation requests 16 tokens. The first five generated IDs are [12095, 13, 576, 6722, 315]. At zero-based generated position 5, the committed our_ids.npy expects 9625 and the CPU emits 15344. The boundary case fails on prompt 0 before it reaches its separate boundary exercise.

A fresh baseline build and an independent operator rerun reproduced the same two failures. Both report 28 passing assertions and two failures. The baseline worktree is clean, and the tested executable has SHA256 46c8a80062219b0135b25ef0668de9b79232fd8f1cf6c173fe4194d8fa6aaf9d. The F16 implementation's full CTest run encounters the same failure. This establishes that the failure predates that implementation. It does not establish whether the product, artifact selection, or anchor is wrong.

The owning spec is .agents/specs/first-additive-model-qwen3-dense.md. The model row records earlier France/Italy differences for Metal and a portable CUDA build. Those observations do not adjudicate this CPU execution. Issue #2671 concerns dump-mode verdicts and does not resolve this ordinary non-dump run. Issue #269 concerns a different gfx1200 ROCm failure.

Evidence in the supplied local workspace is under /home/vikash/vllm.cpp-rdna3-f16-impl/build-rocm-f16-evidence/: cpu-paged-baseline-v1-command.json, cpu-paged-baseline-v1-operator.log, and cpu-paged-baseline-v1-operator-receipt.json. The command and receipt record the executable, source revision, environment, and exit status.

The owning model row owes an active-pin, artifact-matched diagnosis and a reviewed correction. Preserve the current fixture, anchor assertion, and tolerance while diagnosing it. Do not refresh the golden solely to make this run green. Keep this full-suite result failing until the cause and correction pass the declared gate.

Start with tests/parity/test_qwen3_paged_engine.cpp:365 and the CPU test command recorded in build-rocm-f16-evidence/cpu-paged-baseline-v1-command.json. Compare the ordinary and KV-boundary failures with the owning spec and operator receipt, checking artifact and active-pin selection before changing anything. Done means a reviewed cause and correction pass the declared gate while preserving the fixture, anchor assertion, and tolerance.