Layer streaming: a bad SOUP_LAYER_STREAM_STRIPE_DIRS entry is refused only after the dataset loads; --dry-run and soup doctor never check it #1613

Follow-up to the layer-cache striping merged in #1536. Line numbers are from main at the merge
commit, 45b8470c.

What happens today

The entries of SOUP_LAYER_STREAM_STRIPE_DIRS are parsed and validated in one place: inside
_setup_streaming_transformers, after the tokenizer and the model config have been loaded
(src/soup_cli/trainer/stream_setup.py:636-645, then resolve_stripe_roots at :673-678).
That method runs from trainer_wrapper.setup(dataset), which is late in soup train
(src/soup_cli/commands/train.py):

Observed in a real run on Windows, on the #1536 branch before it merged (SmolLM2-135M-Instruct,
an 8-row dataset, the variable set to a folder that does not exist): exit code 1 after about
21 s, after the dataset had loaded, a Run ID had been printed and the tokenizer had been
loaded. The run tracker kept a failed row for it. The output was:

Error: StripeRootError: SOUP_LAYER_STREAM_STRIPE_DIRS entry
'D:\\does-not-exist-xyz': must be an existing directory. Soup creates only its
per-model folder inside a stripe root, never the root itself, so a drive that
is not mounted cannot be mistaken for an empty one. Reconnect the drive, or
remove this entry (unset SOUP_LAYER_STREAM_STRIPE_DIRS to re-shard to one
root).
Run with --verbose for the full traceback.

Three more things follow from the same placement:

• --dry-run cannot report the problem. Its branch exits before a trainer exists, so it never
  reaches the check. (Read from the code; a dry run with a bad entry was not executed.)
• soup doctor does not read the variable. soup doctor --disk classifies the current
  directory only (src/soup_cli/commands/doctor.py:948-960).
• The message starts with the exception class name, because format_friendly_error has no
  entry for it and falls back to <type>: <message> (src/soup_cli/utils/errors.py:364).

What should happen

• An entry that breaks a rule which needs no disk probe is refused before the dataset is loaded
  and before a run row is created.
• soup train --dry-run reports the same refusal.
• soup doctor shows the variable and what Soup makes of each entry.
• The refusal reads as a refusal with a way out, not as StripeRootError: ....

Suggested direction

• When training.stream_layers is on and the variable is set, run the cheap rules right after
  the config is loaded: parse_stripe_dirs, the count cap, validate_stripe_root for each
  entry, and the same-volume rule (volume_of, one os.stat). utils/stripe_roots.py has no
  top-level torch import, so the CLI stays light if the import is inside the function.
• The rules need the primary cache ROOT, not the per-model folder. resolve_shard_dir creates
  the per-model folder as a side effect (src/soup_cli/utils/layer_shard.py:696-709); an early
  check should take the root without creating anything.
• Leave the NVMe rule where it is: the disk-kind probe costs about 9 s on Windows. Keep the full
  validation in the setup as the authority; the early call only moves the common mistakes
  forward.
• Call the same helper from the --dry-run branch.
• In soup doctor, add a row per entry. --disk already pays for the probe, so with that flag
  the NVMe rule can run too.
• Give StripeRootError a friendly form in utils/errors.py, or catch it in the command.

Acceptance test

• A CliRunner test of soup train with stream_layers: true and the variable naming a folder
  that does not exist: exit code 1, the output contains "must be an existing directory", it does
  not contain the "Loaded:" line or a Run ID, and ExperimentTracker.start_run was not called.
• The same config with --dry-run: the same refusal, and no "Config valid" line.
• A same-volume entry is refused at the same early point.
• soup doctor with the variable set to a missing folder prints a row that names the entry and
  the rule; with the variable unset the output is unchanged.
• With the variable unset, soup train behaves exactly as before (no new output, no new call).
• Output assertions go through the ANSI-strip and whitespace-collapse helper the test suite
  already uses, and the tests pass with FORCE_COLOR=1.
• Mutation check: remove the early call, and the first two tests fail.