fix(BACKEND-TENSTORRENT): route the exact-f32 decode dots through mean·K — the tt-metal pin moved f32 ttnn::sum onto the tf32 FPU tile-reduce

[lu-zero]

Two of the three P150 device-test failures flagged during the #3475 review share one root cause and are fixed here; the third is the pre-recorded in-suite-only RAC flake and is left to its owed bisect.

Root cause

The exact-f32 dots in the composed GDN decode step and the grouped keep-quant P==1 decode are broadcast multiply (exact f32 SFPU) + ttnn::sum over the K dim. The tt-metal stack this backend builds against since 252ca04 (tt_umd 0.9.12 era, /tmp/umdtrial-install) moved f32 ttnn::sum onto the tf32 FPU tile-reduce — the pin's own generic_reductions.hpp documents the split: mean's default fast_and_approximate_mode=false is "the accurate SFPU path (full fp32)", true is "the faster tf32 FPU path", and sum has no SFPU arm. Neither GDN nor KEEPQUANT product code changed after the 2026-10-04 102/103 suite run, so the pin advance is what broke the pins.

Red-first evidence (P150, deterministic, 3/3 identical runs each)

• kGdnDecode wide-range arm: state max_abs 0.0051, max_rel 1.83 vs the 0.002/1e-5 envelope (test_tenstorrent_backend.cpp:4126). The error lands in the small-magnitude delta-rule terms — the exact signature 8ab35b9 pinned the arm against. VT_TT_GDN_DECODE=chunked degrades further (max_abs 0.246), ruling out a composed-only defect.
• kMatmulBTQuantGrouped decode P=1 Q4_K N=1024 K=5120: bad=154/1024, worst_rel 0.1307 (:9696), bit-identical across the fused arm, VT_TT_KEEPQUANT_FUSED=0, and VT_TT_KEEPQUANT_MM_CHAIN=1 — the shared ttnn::sum is the defective stage, not a matmul arm.

Fix

Reduce through ttnn::mean(dim, keep_dim) and rescale by the exact f32 K (two ulp of rounding against tf32's 2^-11 truncation) at the three sites. No assertion re-pinned, no envelope widened.

Green evidence

• GDN wide-range: state rel_rms 3.6e-5 → 3.2e-8, max_rel 1.83 → 9.1e-4.
• Grouped decode: worst_rel 0.1307 → 1.7e-4, bad 0/1024.
• Full suite on the P150 (2026-10-19, default order): 107/108 cases, 791117/791119 assertions. The only failure is the batched decode RAC capture-safety case (:2550/:2551), which passes isolated 3/3 and whose in-suite failure values vary run to run (126/128 subtle drift on the pre-fix baseline, 106/128 with inf after): the pre-recorded 2261-class RAC flake whose bisect is an owed item (.agents/specs/tenstorrent-decode-fusion.md:241,300), not touched here. The post-summary teardown segfault (exit 139) is the known defect recorded in 0694cdf.
• test_kolibri1_tt host suite: 17/17 cases, 246/246 assertions.
• agent-preflight.sh --staged, check-agent-record.py, commit style and trailer gates: all green.

Issues

• ISSUE-LOCAL-01M590MZYN28EA6DEKFFA8GNVY (BACKEND-TENSTORRENT-GDN)
• ISSUE-LOCAL-01M590NESYZ6D3QNF3F34DDXJG (BACKEND-TENSTORRENT-KEEPQUANT)

Filed and closed with this fix's evidence.

FOLLOWING_AGENTS_PROTOCOL

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