[PTO 0.58.2][MC/LLD] Repin bc369ec projection and implement CUBE Layout/B.FPATR contracts

[zhoubot]

Summary

The currently tracked LLVM integration commits do not identify or implement the latest PTO 0.58.2 encoding projection at pto-spec@bc369ec67a07c0260f6ba793fa0d705abb363770.

Current downstream references:

• specified integration branch: temp/shared-tload-integration-20260811@eb64de8afcbda043aec7e56dae346905dc982039
• active branch: dev-llvm15_56@b945a5d01323de2f38989288d0a4c423abff3ad1

Both precede the PTO CUBE contract series merged on 2026-08-20/21.

New PTO identity and encoded surface

release                    = 0.58.2
encoding_abi               = pto-isa-0.58.2-mode-function-v1
tag v0.58.2 projection     = 9738fbfb67c90b90dacdae926c6ee206a2d517691ae27e5e5fc019702cc0e447
current main projection    = 2f94ed85e15f751131306fafbb2dcf2dcf05227ca5a217adb8c8093b5d5a5957

The current PTO head additionally assigns:

• B.DATR.Layout 21–26: ND2M32/ND2M16/ND2N8/M322ND/M162ND/N82ND
• B.FPATR bits 7/8: TransA/TransB
• byte-valued B.IOR row stride for TLOAD/TSTORE
• persistent Local CUBE M16/M32/N8 state
• explicit accumulator C with distinct D publication for ACC Matrix forms

The available local Linx compiler used by downstream smoke currently emits a different PTO note:

{"encoding_abi":"pto-isa-0.58.0-mode-function-v1",
 "encoding_projection_sha256":"0cad2272ada8f53fc8354e22568099fe8d6bd4b7832c837260cd370b0fc76ffa",
 "release":"0.58.0"}

That binary is diagnostic evidence only, but it demonstrates that no current, consumable compiler tuple is available through the tracked build flow.

Requested LLVM/LLD closure

1. Pin the exact PTO source/projection used by MC and LLD.
2. Emit .note.pto.isa for projection 2f94ed85... and reject old/mixed projection objects.
3. Add parser/printer/encoding tests for Layout 21–26 and TransA/TransB.
4. Preserve exact TLOAD/TSTORE byte-stride operand order and unit.
5. Cover distinct C/D operand roles for TMATMUL.ACC, TMATMULMX.ACC, TGEMV.ACC and TGEMVMX.ACC.
6. Validate both llvm-mc and Clang integrated assembler surfaces.
7. Publish the merged LLVM SHA for toolchain-build to pin; do not leave release evidence on a topic-head-only compiler.

Related canonicalization tracking: #38.