Transcription temperature: malformed values are silently reinterpreted instead of rejected

Describe the bug

/v1/audio/transcriptions silently reinterprets a malformed temperature field instead of rejecting it, because a second, looser parser rescues values the first one correctly refused.

src/audio/speech_to_text/s2t_servable.cpp:75-90:

auto temp = ovms::stof(temperatureStr);
if (!temp.has_value()) {
    temp = ovms::stou32(temperatureStr);
    if (!temp.has_value())
        return absl::InvalidArgumentError("Invalid temperature type.");
}
config.temperature = temp.value();
if (config.temperature != 0) {
    config.do_sample = true;
}

ovms::stof (src/stringutils.cpp:196-216) requires the whole string to be consumed and rejects NaN/Inf. ovms::stou32 (src/stringutils.cpp:112-129) does neither — it is the only converter in stringutils that does not check idx against the string length, and it calls erase_spaces() on its input first.

So the fallback accepts exactly the values stof had just rejected:

field	parsed as	correct result
temperature=0.5x	0.0 — not 0.5	400
temperature=0abc	0.0	400
temperature=1e400	1.0 (stof throws out_of_range, stou32 parses the leading 1)	400
temperature=5 5	55.0 (spaces erased first)	400

Verified by compiling both versions of the function with ovms::stof, ovms::stou32 and erase_spaces copied verbatim:

  0.5x        before: accepted t=0 do_sample=false        after: 400 Invalid temperature type.
  0abc        before: accepted t=0 do_sample=false        after: 400 Invalid temperature type.
  1e400       before: accepted t=1 do_sample=true         after: 400 Invalid temperature type.
  5 5         before: accepted t=55 do_sample=true        after: 400 Invalid temperature type.
  abc         before: 400 Invalid temperature type.       after: 400 Invalid temperature type.

The 0.5x case is the most user-visible: a typo in a value that looks almost right silently turns sampling off and transcribes greedily, with a 200 response and no warning.

The fallback also appears to serve no purpose. ovms::stof already parses integer spellings ("1" → 1.0f), so an integer-only parser is never needed as a second chance — every input it can rescue is malformed by definition.

Not reporting negative temperature as a bug. temperature=-1 is accepted and enables sampling, but that is deliberate: #4201 ("Negative temperature should lead to genai error in STT path") established that such values are passed through for GenAI to reject, and negativeTemperatureEnableSampling in src/test/audio/speech2text_test.cpp pins it. Noting it here only so it is clear it was considered and intentionally left alone.

To Reproduce

curl http://localhost:8000/v1/audio/transcriptions \
  -F file=@audio.wav -F model=whisper-tiny -F temperature=0.5x
# 200 OK, transcribed with temperature 0.0 (sampling disabled) - expected 400

curl http://localhost:8000/v1/audio/transcriptions \
  -F file=@audio.wav -F model=whisper-tiny -F temperature=1e400
# 200 OK, transcribed with temperature 1.0 - expected 400

Expected behavior

A temperature field that is not entirely a number returns 400 Invalid temperature type. — which is what the code already intends; the fallback just prevents it from firing.

Logs

At --log_level TRACE, Received temperature: 0.5x is logged, followed by a normal transcription.

Configuration

1. OVMS version: main @ fadb3314
2. Any speech-to-text servable (e.g. whisper-tiny)
3. CPU
4. N/A
5. Any audio file

Additional context

Suggested fix: drop the ovms::stou32 fallback so the existing Invalid temperature type. error path fires. The three existing SttServableParseTemperatureTest cases (-1.0, 0, 1.0) all parse cleanly through ovms::stof and never reached the fallback, so they are unaffected.

Related: the root cause is that ovms::stou32 does not verify the whole string was consumed, unlike stou64, stoi32 and stof in the same file. That is filed separately as a stringutils inconsistency, since it also affects the GRPC_SERVERS environment variable. Fixing either one independently fixes this endpoint; the two do not conflict.

I have a patch for this and will open a PR shortly.