Integrate the bfp16ebs8 vendor experiment: XBFP container and on-metal suite (#148)

[aravishankar-mp]

Why

Issue #148 crate integration (design: docs/plans/issue-148-bfp16ebs8-vendor-tier.md, on PR
#161's branch): the AMD bfp16ebs8 vendor experiment's XBFP container, its load_program
arm, the vendored #146 reference model as a host-runnable test target, and an on-metal suite in
its own test file. Draft until two gates clear: PR #161 (which carries the probe pipeline
that builds the fixture and the docs this references) and the P6 question below.

The container carries only shape and payload; the slot plan is derived, never self-declared.
Loading validates the experiment's own immutable target identity — presenting any other
identity (including a TOSA target) is a relabeling attempt and is rejected — then translates
into the internal precompiled form so executable construction and submission reuse the one
audited path. Nothing is advertised: no TosaCapabilityProvider row, no TOSA format
acceptance, no protocol value.

Proven on the reference NPU (fail-loud lane, 1022:17f0): the full guest-quantized MXINT8
MATMUL lifecycle at K = 512 is bit-exact against the fold oracle; foreign identities are
rejected at load; the fixture's derivation is pinned. The vendored model's nine silicon-fixture
tests also run on every host without an NPU.

The important finding is a falsification. The planned accumulation contract — per-step FP32
round-to-nearest-even in ascending-k order — is wrong at tie-adjacent steps. Probe P6
(tests/bfp_p6_probe.rs, manual on-metal measurements, findings in its module doc) shows the
mac chain is serial with no guard bits persisting across an X, d, −X sequence, yet crafted
exact ties break toward zero while an organic mid-chain tie broke away from zero — refuting
RNE, floor, to-odd, half-away, half-toward-zero, and every wider-precision single-rounding
model. Until P6 pins the real rule, the order-contract hardware test is quarantined behind an
#[ignore] naming the reason, and the tier claims bit-exactness only away from exact ties.
Shipping the falsified oracle as a green test would have been worse than shipping no test.

Compatibility

• No wire effect. This changes no accepted or emitted protocol bytes. The format
  constant is backend-local and the experiment is reachable only by presenting its own
  identity.

Checklist

• Does not alter payload lengths, ownership, reset, error, timeout, or feature-negotiation
  behavior.
• Authoritative inputs not regenerated.
• Public Rust API additions called out: bfp_experiment module (format constant, identity,
  parse/encode). Existing APIs unchanged.
• No dependency changes; no platform behavior moved into a portable crate.
• No unsafe code added.
• The experiment remains unadvertised and documented as out of scope for TOSA capability.

Verification

cargo fmt --all -- --check
cargo clippy -p virtio-accel-xdna --all-targets --all-features --no-deps -- -D warnings
cargo test --workspace --all-targets
VIRTIO_ACCEL_HRX_LIB_DIR=/tmp/fakelib cargo clippy -p virtio-accel-xdna --all-targets --all-features --no-deps -- -D warnings

On metal (v2026.08 toolchain, HRX 2026.07.30):

VIRTIO_ACCEL_XDNA_REQUIRE_HARDWARE=1 cargo test -p virtio-accel-xdna --test bfp_experiment -- --test-threads=1
# 12 passed; 1 ignored (the quarantined order-contract test)
cargo test -p virtio-accel-xdna --test bfp_p6_probe -- --ignored --nocapture
# manual P6 measurements