[FU-max-boop]

Summary

• retain every multiprocessing.AsyncResult and inspect it after all DUT workers finish
• fail the parent process when any supported DUT build fails
• distinguish six intentionally unimplemented build scripts with UnsupportedDUTError, preserving wildcard and subtree builds
• add hardware-independent regression coverage for worker outcomes, result transport failures, deterministic aggregation, unsupported builders, and prebuilt DUTs

Problem

build_dut() submitted _build_dut() through Pool.apply_async() but discarded every returned result. _build_dut() also only logged falsey returns and exceptions. A supported DUT build could therefore fail while the parent returned normally and the shell reported status 0.

A process-level probe using a real Linux multiprocessing.Pool and a synthetic supported builder returning False reproduced the behavior:

e887def (base): logs the build failure, shell exit 0
31aaffb (head): raises "Failed to build DUTs: build_ut_audit", shell exit 1

Implementation

• _build_dut() returns True for success, False for a falsey result or ordinary build exception, and None only for an explicit UnsupportedDUTError.
• build_dut() records results under deterministically sorted DUT names, waits for all workers, then calls AsyncResult.get() for every submitted build.
• Supported failures are aggregated into one RuntimeError; exceptions from AsyncResult.get() propagate.
• The six existing placeholder scripts explicitly raise UnsupportedDUTError, so broad targets continue to skip them.

Validation

python -m pytest -p no:cacheprovider -q tests/test_build_dut.py

• Python 3.10.20 / pytest 8.3.4: 16 passed
• Python 3.12.3 / pytest 9.1.1: 16 passed
• final regression file applied to base e887def: 14 failed, 2 passed
• real multiprocessing process probe: shell exit 0 -> 1
• real comm re-export plus all six placeholder imports: passed
• git diff --check: passed

Related work and merge ordering

#196 propagates pytest exit codes later in run.py; this change covers the earlier DUT-build stage and does not overlap its files.

A follow-up RAS contribution replaces scripts/build_ut_frontend_bpu_ras.py with a real builder. This PR should land first; the RAS branch will then retain its builder and drop that script from the unsupported set during rebase.

Scope boundary

• Targeted tests do not compile RTL or invoke Picker.
• The process-level probe uses a real multiprocessing pool with a synthetic failing build function, not a full hardware compilation.