fix: preserve negative sign when parsing float fields in TrainingConverter

_FLOAT_PATTERN had no `-?`, so _parse_float('-5.2') returned 5.2 and
"-3 degrees" parsed as 3.0 -- silently flipping the sign of any negative
value during TrainingConverter's field-by-field fallback extraction
(used when a smaller LLM's direct-to-Pydantic conversion fails).
Reachable for any Pydantic model field typed `float` via
_process_field_value -> _parse_float.

Add the missing `-?` to the pattern.

AI-Generated PR: Yes, drafted with Claude Code and personally
reviewed/tested before submission. Per CONTRIBUTING.md's llm-generated
label requirement: as an external contributor I cannot self-apply GitHub
labels (confirmed via `gh pr edit --add-label`, GraphQL permission
error) -- disclosing here in the PR body/commit instead; happy to have
a maintainer apply the label if needed.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: chuenchen309

lib/crewai/src/crewai/utilities/training_converter.py

@@ -7,7 +7,7 @@
 from crewai.utilities.converter import Converter, ConverterError
 
 
-_FLOAT_PATTERN: Final[re.Pattern[str]] = re.compile(r"(\d+(?:\.\d+)?)")
+_FLOAT_PATTERN: Final[re.Pattern[str]] = re.compile(r"(-?\d+(?:\.\d+)?)")
 
 
 class TrainingConverter(Converter):

lib/crewai/tests/utilities/test_training_converter.py

@@ -95,3 +95,12 @@ def test_process_field_value_float(self):
         response = "The quality score is 8.5 out of 10"
         result = self.converter._process_field_value(response, float)
         assert result == 8.5
+
+    def test_process_field_value_float_negative(self):
+        response = "The temperature delta is -3.2 degrees"
+        result = self.converter._process_field_value(response, float)
+        assert result == -3.2
+
+    def test_parse_float_preserves_negative_sign(self):
+        assert TrainingConverter._parse_float("-5.2") == -5.2
+        assert TrainingConverter._parse_float("-3 degrees") == -3.0