record(LTX25-DECODE-SPEED): three scalar single-threaded loops in the LTX CPU path, and the number ltx2_video.cpp said was owed

[localai-bot]

record(LTX25-DECODE-SPEED): three scalar single-threaded loops in the LTX CPU path, and the number ltx2_video.cpp said was owed

Three defects found while attempting the first full-model LTX-2.5 render are
filed and now have an owner. None is fixed here; this change is the record.

They are one story. The LTX prep path hand-rolls scalar arithmetic where the
rest of the tree routes through the vt:: GEMM seam, and it does so in three
places that all sit between Load and the first denoise step:

• Ltx2FuseLoraIntoTensor is a scalar single-threaded loop (~0.53 GFLOP/s), blocking every LoRA arm on the full 21B DiT #1202 Ltx2FuseLoraIntoTensor computes the LoRA product with a triple
  scalar loop at a measured ~0.53 GFLOP/s, one thread of twenty.
• LTX-2.5 text-tower Linear is a scalar single-threaded double-accumulating loop; dominates pre-generation wall and reads as a hang #1208 the text tower's Linear is scalar, single-threaded and accumulates
  in double, which also diverges from the F.linear it cites as reference.
• The two-stage LoRA rebind cost is measured: load-fuse is immediately undone by phase 0, and each pass is hours at #1202's rate #1210 the two-stage rebind fuses at load, un-fuses for phase 0 and re-fuses
  for phase 1, so the load-time pass is provably wasted.

#1210 supplies a number the tree explicitly asked for. ltx2_video.cpp:2849
records the two-rebind cost as "UNMEASURED on real weights" and says a later
perf row owns it. This is that measurement.

The owner is ltx25-decode-speed.md, whose §5 already frames "why the decode
is single-threaded and on the host". These extend that finding from the decode
to the load and prep path, so they belong to its ## Owed table rather than to
a new row.

Scope is deliberately narrow. The rates are measured over a 10.4-minute window
and extrapolated; no full-model pass has completed, so no end-to-end figure is
claimed. What is measured is stated as measured and what is projected is
labelled.

test_cpu_x86_llamacpp_floor fails in the local preflight with exit 4
(NO_QUIET_WINDOW) instead of 2 at loadavg 68.9-79.9, which is #618 exactly.
This diff touches two files under .agents/ and cannot reach a CPU benchmark
harness. check-agent-record.py, check-public-doc-tables.py, doc-checkpoint
and now-current all pass.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [claude-code]