Context cleanup removes the timestamps needed to recognize chat conversations

## Problem

The timestamp-based conversation detector works on raw screen fixtures but loses its evidence on the production context path. `Surroundings.trimmed` removes timestamp parts before `Register.infer` sees them. In a chat whose speaker names and timestamps occupy separate elements, neither timestamp-based turns nor colon-prefixed speaker turns remain. The register treats the reply as an ordinary line and omits its conversation hint.

Reviewed and reproduced on main `8d93f5b8574f3c7b8c30086d9c817bf7f5d4827d`.

## Evidence

- [Collector cleanup](https://github.com/uttrflow/uttrflow-swift/blob/8d93f5b8574f3c7b8c30086d9c817bf7f5d4827d/Sources/UttrflowContext/Surroundings.swift#L186-L194) calls `Timestamps.without` on each element.
- [Coordinator](https://github.com/uttrflow/uttrflow-swift/blob/8d93f5b8574f3c7b8c30086d9c817bf7f5d4827d/Sources/Uttrflow/Suggestion/SuggestionCoordinator.swift#L555-L575) passes `around.text` as generation surroundings.
- [Register](https://github.com/uttrflow/uttrflow-swift/blob/8d93f5b8574f3c7b8c30086d9c817bf7f5d4827d/Sources/UttrflowPredict/Register.swift#L155-L175) recognizes timed turns by finding at least two clock-bearing lines beside a message composer.

An isolated test uses the real `Surroundings.collect` with the repository's fake element tree. Its window contains separate static-text elements:

```text
Chats
Neha
10:31 AM
Standup moved, please confirm.
Arjun
10:33 AM
Confirmed.
```

The focused field is a message composer (`Message #platform`). The collected text is:

```text
Chats
Neha
Standup moved, please confirm.
Arjun
Confirmed.
```

Passing the original lines to `Register.infer` gives `isConversational == true`; passing the actual collected text gives `false`. The expected conversation assertion fails. This is the same layout covered by the existing `RegisterConversationTests` team-chat fixture, but those tests bypass the collector. All 38 selected original tests in this review batch pass; the added boundary test fails.

No live Accessibility session or model-generated reply was measured, so this reports the verified classification and prompt-hint defect, not a measured reply-quality rate.

## Acceptance

Preserve conversation evidence across collection and inference while still keeping timestamps out of generated prose. For example, carry structured turn evidence separately from cleaned prompt text. Add a collector-to-register regression for the separate-name/time layout and the inline timestamp layout already represented by the register tests. Retain the negative page-navigation cases from #454.

#454 addresses false positives on non-chat screens; this is a separate false negative introduced by cleanup before inference. Checked open and closed issues and current open PRs before filing.

Start with the cleanup in Sources/UttrflowContext/Surroundings.swift, then trace SuggestionCoordinator.swift into Register.swift and compare the collector path with RegisterConversationTests. Run the focused collector-to-register regression and the existing conversation and #454 negative cases; done means both separate-name/time and inline-timestamp layouts retain conversation detection without putting timestamps into generated prose.

Written by the indexing model from the issue text.