[do-not-mirror][do-not-merge] chore(fork): restore the shadowed round-four regression tests - #21

The Linux CI jobs fail while installing the package:

    [Errno 122] Disk quota exceeded

Errno 122 is EDQUOT, hit while unpacking the CUDA torch stack that
`pip install -e .` resolves by default. torch alone is 554.6 MB and the
accompanying nvidia-* wheels add another ~2.4 GB, which no longer fits
the runner's quota. The failure occurs during "Installing build
dependencies", i.e. the isolated PEP 517 build environment is already
enough to exceed it on its own.

The trigger is a runner image change, not a code change:

  * upstream 11_bernhard_dev, 2026-10-20, Image: ubuntu-24.04 -> success
  * same workflows, 2026-10-23, Image: ubuntu-26.04 -> failure

Both resolve an identical wheel set (torch 2.14.1, 554.6 MB). The
`ubuntu-latest` label began migrating to Ubuntu 26 on 2026-10-19
(actions/runner-images#14748), between those two runs. Both workflows
already carry "Free up disk space" steps, so these jobs were running
close to the limit before the migration.

Fix: install CPU-only torch from download.pytorch.org, then build with
--no-build-isolation so the build reuses it instead of resolving torch
a second time. --no-build-isolation needs the build dependencies present
up front, hence the explicit setuptools/wheel/numpy install; the
`setuptools<81` bound matches the pin already in install_requires.

This loses no coverage: BUILD_CUDA_EXT keys off CUDA_HOME/ROCM_HOME,
which is unset on these runners, so the CUDA sources were never compiled
here. tests/smoke_tests/Dockerfile already installs torch this way.

macOS keeps the plain `pip install -e .` unchanged: its PyPI wheels are
already CPU-only, and leaving it alone keeps the default PEP 517
build-isolation path under test somewhere in the matrix.

The shipped puffer_drive.yaml sets policy.fused_slot_encoder=true, and
DriveBackbone raises ImportError when that is set without triton. Two CI
jobs build a policy straight from the shipped config, so Unit Tests and
Notebook Execution have been failing on every pull request:

  ImportError: fused_slot_encoder=true requires triton

The smoke and eval suites already pin the flag off for exactly this reason;
these two call sites were simply missed. test_action_selection only exercises
the discrete-to-continuous action table, which the fused kernel has no part
in, so it pins the flag off outright. The inference notebook keeps the kernel
whenever it has a CUDA device and drops it otherwise, which is also what a
reader running the notebook on a laptop needs.

Neither change touches the training path or the kernel itself: with triton
and a GPU present, both call sites behave exactly as before.

The two instruction files are byte-identical except for one line, and that line is a contradiction rather than a Claude-specific tweak. AGENTS.md asked for comments that explain invariants; CLAUDE.md banned them by default with a hard one-line cap and an explicit list of banned forms.

Agents load one file or the other depending on the tool, so the same codebase was being written to two different standards, and neither copy said the other existed. This adopts the stricter rule in both, because it is the one written for this codebase and the one the recent code follows.

The git-identity fix added a second class literally named
TestRoundFourRegressions. Python rebinds the name, so the original
class object was discarded before unittest could collect it and its
six tests stopped running silently.

Rename the newer class to TestRoundFiveRegressions. The suite collects
and runs all 133 methods again, up from 127.