fix(tests): repair UTF-8 corruption in TrainingContractParityTests doc comment (#1898)

Utf8EncodingTests.SourceFiles_ShouldNotContainReplacementCharacter fails on master:
TrainingContractParityTests.cs carries a U+FFFD replacement character where an em dash
belongs ("...which the facade calls once <U+FFFD> so their callbacks fired...").

Because CI builds each PR MERGED with master, this red check surfaces on every open PR
touching this repo even when the PR has nothing to do with it — it turned up on #1897,
whose diff was confined to src/TimeSeries. Fixing it at the source unblocks all of them.

Restores the intended em dash (U+2014). A repo-wide scan confirms this was the only
U+FFFD occurrence in tracked source. The test the check runs now passes locally.

Note: the failure message advises `python3 scripts/fix-encoding.py`, but that script does
not exist in the repo — worth either adding it or updating the message in a follow-up.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

Author: ooples

tests/AiDotNet.Tests/IntegrationTests/Training/TrainingContractParityTests.cs

@@ -22,7 +22,7 @@ namespace AiDotNet.Tests.IntegrationTests.Training;
 /// <remarks>
 /// Neural networks reach training through the optimizer path, which already drives the facade's
 /// per-epoch seam (OptimizerBase.SetEpochProgressCallback -> InvokeTrainingEpoch). TimeSeries models
-/// instead own their epoch loop inside Train(), which the facade calls once � so their callbacks fired
+/// instead own their epoch loop inside Train(), which the facade calls once — so their callbacks fired
 /// a single synthetic epoch after training had finished, and no veto could take effect. These tests
 /// assert the neural side genuinely works, so that the two families cannot silently diverge again.
 /// </remarks>