fix(agent): a null choice label made the answer undecodable

[PerryLink]

What this fixes

A null choice label is accepted today, and the answer that comes back cannot be decoded.
67edd5f rejected a null score level on the grounds that a label "is rendered as option text
and used as the answer key" — a choice label is used for exactly those two things, and the check
was not extended to it. Measured on main (4066d5d), English checkpoint, through system_one:

q = {"q": {"type": "choice", "instructions": "which?", "criteria": ["billing", None]}}
agent.system_one("my invoice is wrong", q)["answers"]["q"]

{"type": "choice", "choice": "billing",
 "probabilities": {"billing": 0.7679, "null": 0.2321},
 "confidence": 0.2183, "answer_confidence": 0.7679}

Two problems in that one response.

The option text and the answer key disagree. _to_internal normalises a list of labels to
{label: None}, so render_options renders the label through str(k): the option the model chose
between was billing and the text None. The answer key and the probabilities key for that
same option are "null", because a dict key has to be a str.

"null" is not decodable. A client cannot tell whether that option was the string "null" or
JSON null, because both produce the same key:

criteria = ["billing", None]
criteria[answer["choice"]]   # -> None, but the key said "null"
answer["choice"] == "null"   # -> False, so the key is unreachable from the response

So an answer that includes the null label is not usable by a client that round-trips the key back
into the criteria it sent, which is the documented pattern.

After

{"criteria": ["billing", None]}
-> ValueError: question 'q': choice label 1 is null; a label is rendered as option text and used
   as the answer key, so it must be a string, number or bool -- a null label renders as the text
   "None" while its answer key is "null"

Over /v1/systemone that is a 422, since serve maps ValueError from the agent to 422.

The one judgement call worth your opinion

This is a behaviour change: it turns a 200 into a 422 for anyone sending a null label today.
The alternative is to keep accepting null and make the round-trip work instead, by rendering the
option text as "null" so it matches the key. Both are defensible, and I picked rejection only
because 67edd5f rejected the score-level case rather than stringifying it — if you would rather
have the other, this is a two-line change.

"" is deliberately not rejected: unlike a null it already round-trips (render_options
renders it as "", the answer key is "", and criteria[""] finds it), and it is pinned in the
tests so the fix is not generalised to "reject the empty-looking labels".

Scope

• Dict-shaped criteria is untouched. A None value there is the documented "no description" spelling used throughout tests/, and a dict key cannot be None anyway.
• One guard, both backends. It goes in Agent._check_question, which ONNXAgent._encode_states already calls (onnx_agent.py:285), so the two cannot drift the way the lang_temperatures parsing did.
• A nested label is a different error. The existing guard from fix(serve): a request with no state was answered about the literal text null #427 still fires first for a list/dict/set label.

Relationship to the existing checks

laya.structured already rejected this shape one layer up, in different words:

test_structured.py:127   [None, "null"] -> SchemaError: enum values produce duplicate choice labels

The collision was recognised for JSON schemas but not for a directly-written question. This closes
the direct path and leaves the schema path as it is.

@mandu5's #302 reports this area. Its point 2 (a null score level) was already fixed by
67edd5f before this PR — I have said so on the issue
rather than open a duplicate PR for it — and this is the choice half, which that commit left open.

Checks

python tests/test_criteria.py            149 passed, 0 failed   (139 before)
python tests/test_router.py              546 passed, 0 failed
python tests/test_batch.py                49 passed, 0 failed
python tests/test_structured.py          all structured tests passed
python tests/test_onnx_lang_parity.py     11 passed, 0 failed
python -m pytest tests/test_serve.py tests/test_router_batch.py \
                 tests/test_system_one_lang.py tests/test_audit_regressions.py -q
65 passed
python -m ruff check laya/ --select=E9,F63,F7,F82,F401,F811 --line-length=120
All checks passed!
python -m compileall -q laya/ tests/

All of these run in CI, and the whole set is weight-free — the tests use the existing
_tiny_agent() in tests/test_criteria.py, not a downloaded checkpoint. A test (python 3.14)
run is not possible for me locally, so the CI matrix is the check for that.

[NandhaKishorM]

Thanks @PerryLink. Merged, combined with #496 in _check_question: a null label is refused first, then repeated or unhashable ones. It ships in 0.3.21.