feat: name the classes a pre-label run will and will not ask for

[JArmandoAnaya]

A pre-labeling run's prompt is the batch's pinned schema narrowed to the classes a bare box
prediction can be written as. A class is dropped for either of two reasons — it does not admit
bbox, so a detection has no shape to land as, or it declares a required attribute, which a
model's answer carries no value for — and until now both exclusions were silent everywhere.
Somebody with a vehicle class requiring a color started a run, watched it complete, and found
no vehicles labeled, with nothing anywhere saying why. That is the gap this closes (#677).

prompt_plan in visionset/inference/prelabel.py now returns both halves of the split, and
detectable_classes reads its asked list rather than computing its own, so the prompt a run uses
and the plan a surface publishes cannot be two derivations. reasons is a list because both
reasons can hold against one class: a caller told only that a class admits no box would add one
and watch it stay absent from the next run's prompt.

GET /batches/{batch_id}/pre-label publishes the plan before a run starts. It guards on the
launch's own terms, so a batch nobody may pre-label and a schema with no askable class refuse
there with the codes the launch already answers, and it takes no connection, since the prompt is
a property of the pinned schema alone. The dialog reads it on open and shows the asked-for classes
with the left-out ones beside them, and shows them again under a settled run's summary, which is
where a run that labeled nothing is actually read. At a terminal, visionset batch pre-label
writes the same two lines to stderr before the first forward pass.

PreLabelExclusionReason is an open vocabulary. It travels as a list a client renders member by
member, so a release that finds a third way a class cannot hold a detection must not cost an older
client the whole plan. The browser drops a reason it cannot word and still names the class, which
is the half that must not be lost.

Found, not fixed

The MCP surface gets nothing here: pre_label_batch still says nothing about the prompt. That
work is on a follow-up branch stacked on this one.

Issue-number references exist in prose in docs/batches.md and docs/mcp.md that predate this
change. They are left alone rather than swept in.

Test plan

Every stage run on the rebased branch, each exiting 0:

uv run pytest tests/architecture tests/cli tests/examples tests/formats tests/inference
    1058 passed, 19 skipped
uv run pytest tests/jobs tests/kernel tests/mcp tests/packaging tests/scripts tests/test_versioning.py
    2104 passed, 5 skipped
uv run pytest tests/server
    791 passed
uv run ruff check .                  All checks passed!
uv run ruff format --check .         419 files already formatted
uv run mypy src/visionset/kernel     Success: no issues found in 73 source files
uv run lint-imports                  Contracts: 4 kept, 0 broken.
pnpm -r build / -r test / -r lint    1062 + 1155 passed, lint clean
bash scripts/check.sh browser        278 passed (41.8s); 1 passed (21.6s)

Nine mutations were run against the new tests to show they fail when the code is wrong: dropping
each exclusion reason, silencing the plan announcement, truncating the route's exclusion list,
hiding the dialog's excluded line, hiding it in done-mode, removing the unknown-reason filter, and
removing the launch guard. Each reddened exactly the tests that claim it.

Closes #677