[lu-zero]

Two keepers from the kolibri P150 residue hunt.

The instrument (cherry-picked from the near-tie row): the B2b-ii gate gains the W3-methodology teacher-forced adjudication — golden prefix, per-divergence gap and top-5 membership, hard-flip counter. "0/8" style verdicts are now "which prompt, which step, how many nats".

The substrate fix: ttnn::rms_norm's rmsnorm_default_compute_config hardwired fp32_acc = false and approx_mode = true — the sum-of-squares accumulated in bf16 even for f32 tiles (which is what vllm.cpp's shadow arm feeds it). With f32 accumulation and accurate rsqrt, the RMSNorm micro moves 0.9899 -> 0.9997 (1 bf16 ULP — the output-store floor). The trial tt-metal tree (/tmp/tt-metal-umdtrial, installed at /tmp/umdtrial-install) carries this on top of the port fixes; the stack-pin record must note it.

Verdicts after: the RMSNorm kernel bias is GONE, but the B2b-ii argmax chain stays 26/33 on both arms — the remaining flips are attributed to the two known op-level drifts (kGdnDecode, kMatmulBTQuantGrouped), which are exactly the TT suite's 2 standing failures. The teacher-forced counts improved (worst gap 5.17 -> 3.19 nats ON) but the hard flips persist: the next localization pass runs on the fixed substrate. The streaming slot-pool throw also still fires on 4 of 8 walks (its message is now just "1" — source unidentified, separate defect).

TT unit suite: only the 2 known failures; W4 EnsureDevice2D and matmul region class split fail only in full-suite order state (pass isolated — order-state flake, recorded).

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:zai/glm-5.3-flash [maki]