Track and pin the TOSA 1.1-draft MX block-scaled contract

[aravishankar-mp]

Part of #110.

Question

How does the official TOSA 1.1-draft MX block-scaled contract change before release, and what exact
released vocabulary would this project eventually need to negotiate?

As of 2026-08-25, EXT-MX-* exists only in the official draft, is marked Experimental, uses
block-32 E8M0-scaled FP4/FP6/FP8/INT8 forms, and does not define “block FP16” or a block-scaled
MATMUL row. Baseline research: docs/research/tosa-int8-block-fp-status.md.

Work

• Pin the official Arm TOSA source revision used for every experiment and record semantic/schema
  diffs as the draft changes.
• Track block-scaled types, scale/value tensor layout, CAST/conversion operators, convolution rows,
  exceptional-value rules, and any future MATMUL contract.
• Do not copy draft discriminants into the stable TOSA 1.0 schema or allocate a provisional public
  extension bit merely to unblock the backend.
• When a suitable specification releases, write the wire-ABI §9 classification and protocol
  adoption plan: explicit negotiation, preserved 1.0 frames, parser/type support, canonical
  vectors, and a new minor-version conformance directory.

Exit criteria

• A release (not merely a draft snapshot) supplies a suitable block-scaled vocabulary, or the
  ticket records that the released contract still cannot represent the intended XDNA operation.
• The exact compatibility/protocol work is decomposed into implementation tickets.
• Standard MX and AMD bfp16ebs8 remain separate names and capability identities unless the
  characterization proves semantic identity.