[Maint] layout_inference golden snapshots are stale on main; harness has no CI hook

maint/layout_inference/run.py currently reports 26 layout drift failures on a pristine main build (verified on main @ 11ec2397 and again on a tree merged with 752312c2 — identical failure set), across the register-count and io-aware snapshots of:

• broadcast_read/rows4_cols1024_t128
• elementwise_copy/fp16_128x128_t128, fp32_64x256_t256
• fp8_copy/128x256_t128, 64x128_t64
• mixed_dtype_chain/128x128_t128, 64x512_t256
• (and further cases; full list from any local run)

The --cute symbolic-vs-oracle checks still pass, so this is golden staleness, not a scorer bug: some layout-affecting change landed after the snapshots were recorded, and since the harness is not wired into CI, the drift went unnoticed. The harness's guarantee has effectively lapsed — new layout work (e.g. #3176, whose own new goldens do pass) cannot use the existing goldens as a regression net.

Suggested fix:

1. Re-record the stale goldens (run.py --record) with human review of the diff, ideally after bisecting which change moved them so the new snapshots are known-good rather than blessed-blind.
2. Add a CI hook for at least the golden check (it is fast; --cute parity could be a second step or stay maintainer-run).

Observed while validating #3176.