fix: preserve reasoning content in streamed chat traces

[Mathew Stevens (mstevens843)]

Related to #2511.

Preserve streamed delta.reasoning_content as message.reasoning_content, concatenating fragments independently per choice. Keep absent, empty and null-only values distinct without letting later nulls erase text. Ordinary content stays separate; reasoning and reasoning_details are unchanged.

Add constructed edge-case regressions and genuine DeepSeek recordings across six pinned/latest OpenAI SDK aliases through both wrapOpenAI and auto-hook instrumentation. Expectations come from ordered raw fragments, with paired snapshots and a patch changeset.

Support the scenario's dependency-bump workflow by forwarding DEEPSEEK_API_KEY to Docker by name and stripping it from dependency-install environments. Two dummy-only behavioral regressions exercise the actual command/environment paths and run in CI.

Fresh validation: 696 unique cases pass, including 87 focused, 589 compatibility, 15 helper/script, one genuine 12-path matrix and four changeset tests. The genuine matrix passes three consecutive keyless replays. Both credential defects were demonstrated RED before correction. Workspace/narrow typings, dependency builds, formatting, lint, script syntax and changeset checks pass.

Earlier identical genuine recordings demonstrated reasoning loss on untouched base after content passed. Optional whole-E2E type checking previously reported 323 identical base/patch diagnostics; it was not rerun in the maintenance pass. Actual Docker execution and other runtime/platform CI lanes remain unrun locally. No new provider requests were made for the maintenance correction.

Credit to the reporter and prior provider work linked in #2511. Prepared and independently reviewed with AI assistance; human review and any personal attestations remain with the submitter.