Self-correction by restatement is asked of the model in the prompt contract and the model does not do it

The clean-up catalogue asks for one thing that nothing in the app actually does. A slot said
twice with a different word after it — "as a gift as a present" — should keep the later one.
The rules deliberately do not attempt it, the model is asked to, and a measurement on
2026-09-06 found the model does not comply. So the words stay, which is the safe side of the
line, but the catalogue currently claims a behaviour the product does not have.

## Where it is

`Sources/UttrflowAI/PromptContract.swift:14` asks for it, in the contract every dictation is
tidied under:

```
- a slot said twice keeps the later: "as a gift as a present"
```

and line 50 shows it as one of the worked examples the model is given.

`Docs/cleanup.md:71` records the measurement and the reasoning:

> **Measured 2026-09-06: the model does not comply.** "I wanted to buy a record as a gift as a
> present" and "let's meet on tuesday on wednesday afternoon" both come back whole.

The same row explains why a deterministic rule is refused, and it is worth reading before
proposing one: the shape "a short frame of function words repeated with a different content
word after each copy" matches "I'll pay for lunch for everyone" and "coffee with milk with
sugar" just as often as it matches a real self-correction. Which one the speaker meant to
stand is semantic. **A pull request that adds that rule will not be merged** — this is settled
and the reasoning is in the doc.

## What to change

The useful work is measurement, not a fix:

1. **Add corpus cases** to `Sources/UttrflowEval/EvaluationCorpus.swift` for restatement, in
the shape of the existing `grammar` block at line 769 — an `EvaluationCase` with `spoken`,
`expected`, `mustKeep` and `mustNotAdd`. Two exist in the doc already and more phrasings
would help: how far apart the two halves can be, whether a trigger phrase ("no, sorry",
"I mean") changes the answer, whether it behaves differently in a message and a document.
2. **Run them and report what the model does**, so this stops being one observation on two
sentences. A case that fails is the point — it makes the gap visible and any future prompt
change measurable.
3. If a *prompt* change moves the number, that is worth a pull request; say what you measured
before and after.

## How to check it

The evaluation harness runs the real pipeline, so this needs an Apple Silicon Mac with the
model installed (`CONTRIBUTING.md` has the setup, and `Docs/eval-methodology.md` explains how a
case is scored). Report the model and the macOS version with any number, and give the exact
sentences — a percentage with no sentences behind it cannot be checked by anybody.

## Notes

- Do not "fix" this by deleting the line from the contract. The instruction is cheap and the
behaviour is wanted; what is missing is evidence about when the model honours it.
- `Docs/cleanup.md:71` is the specification. If the measurement changes the picture, that row
changes with it in the same pull request.

Start with the grammar block around line 769 in Sources/UttrflowEval/EvaluationCorpus.swift, then read CONTRIBUTING.md and Docs/eval-methodology.md before running the evaluation harness on an Apple Silicon Mac. Add restatement cases with spoken, expected, mustKeep, and mustNotAdd, and report exact sentences, model, macOS version, and results; update Docs/cleanup.md:71 if the measurement changes the picture.

Written by the indexing model from the issue text.