Meaning guard accepts moving a negation to another clause, reversing the dictated instruction

## Reproduced behavior

The full `GenerativeTextTransformer` accepts this meaning-changing model response on main `5c310667cd35c44abd8bbd26c19b0905471287b4`:

```text
Dictated: we should not approve the design but we should approve the budget
Returned: We should approve the design but we should not approve the budget.
```

The person forbids approving the design and permits approving the budget. The returned text does the reverse. The words and their counts are unchanged; only the position of `not` changed.

## Verification

An isolated test used the repository's actual `GenerativeTextTransformer`, standard cleaning passes and meaning guard, with `FakeCleanupModel` returning the response above. It called `transform` on a real `TransformationRequest` with an English transcription.

```text
accepted=We should approve the design but we should not approve the budget.
Meaning-changing negation relocation was accepted
```

The rejection expectation fails. This is a controlled model response, not an observed hallucination from Apple's model or a test that inserted text into another application. It establishes that the production acceptance boundary admits the reversal.

34 existing context, number and mirrored-guard tests passed. The mirrored guard tests already reject adding/removing a negation; those count-changing cases do not exercise relocation.

## Cause

In [MeaningPreservationGuard.swift](https://github.com/uttrflow/uttrflow-swift/blob/5c310667cd35c44abd8bbd26c19b0905471287b4/Sources/UttrflowAI/MeaningPreservationGuard.swift):

- `grammarVerdict` sends content words to `survivalVerdict` and `placeVerdict`. Here the content words remain in the same order.
- Negation checks compare total counts in the kept and rewritten token sequences. Both contain one negator, so neither the added nor the dropped check rejects.
- `functionWordChurn` compares multisets. Moving the same `not` contributes zero churn.

Although an alignment now exists, it is not used to enforce where this negation belongs.

## Expected / acceptance criteria

Priority P1: the guard is supposed to prevent a model rewrite from reversing the speaker's instruction.

- Reject a negation moved between clauses while allowing harmless case/punctuation and supported contraction changes.
- Add a full-transformer regression using the example above and its inverse, alongside the existing added/dropped-negation cases.
- Ensure a rejected response reaches the deterministic fallback rather than being inserted as model-cleaned text.
- Evaluate the change against the cleanup corpus before merging; a blanket ban on all token movement would need evidence that it preserves legitimate cleanup.

## Related issues and distinction

Closed #188 addressed added negations. Closed #189 addressed moved content words and permissive stem matching. Open #194 concerns substitutions of prepositions/modals within a churn allowance; this reproduction changes no word or multiset count and bypasses the explicit negation checks. Open #229 is the broader guard-design umbrella; this is a concrete remaining acceptance defect in the current aligned implementation and can be tracked beneath it.