Preserve thinking.display for adaptive-thinking Anthropic requests - #481

)

Adds live payload coverage for #481's preservation of `thinking.display:
"summarized"` on adaptive-thinking requests.

- Capture Sonnet 5 and Opus 5 through Anthropic and Bedrock Anthropic,
including non-streaming, streaming, and follow-up requests/responses (24
native files).
- Capture six transformations to Chat Completions, Responses, and
Google; add 12 request/response snapshot entries. Existing snapshot
entries are unchanged.

Sonnet uses a state-equation problem to exercise adaptive thinking; a
simple arithmetic prompt can be answered without thinking even with
summarized display enabled.

## Capture results

All four native capture suites and all six cross-provider captures
succeeded against real APIs.

| Model / provider | Thinking characters | Streaming thinking characters
|
| --- | ---: | ---: |
| Sonnet 5 / Anthropic | 1,217 | 1,324 |
| Sonnet 5 / Bedrock Anthropic | 1,898 | 2,897 |
| Opus 5 / Anthropic | 16 | 21 |
| Opus 5 / Bedrock Anthropic | 31 | 31 |

## Validation

- `make capture FILTER=AdaptiveThinkingDisplaySummarizedParam`: pass;
force-recaptured Sonnet after choosing the reasoning prompt, then re-ran
capture to confirm artifacts are present.
- `cargo test -p lingua adaptive_thinking_display`: both existing
adapter regression tests pass.
- Reverting only #481's display-preservation logic in an isolated
checkout: both existing adapter regressions fail with `None` versus
`Some(Summarized)`; both pass with the fix present.
- `make test-payloads`: 1,830 pass, 106 skip.
- Payload fixture sync: 853 pass, 74 skip.
- Payload typecheck, focused lint, Rust/TypeScript formatting,
cross-provider guard, and typed-boundary checks: pass.


<!-- sfk:slack-thread -->
[Slack
thread](https://braintrustdata.slack.com/archives/C0AQK98PW5C/p1791476806483969?thread_ts=1791476806.483969&cid=C0AQK98PW5C)

---------

Co-authored-by: Starfolk <noreply@starfolk.ai>