fix(openai): persist complete assistant snapshots

[Yevanchen]

What changed

• publish the authoritative item/completed assistant text as a lossless message.added snapshot
• retain best-effort token deltas for live preview
• update bridge and provider contract coverage

Why

OpenAI token deltas are best-effort. Under uplink backpressure, production retained only the first token of intermediate assistant messages (I, The, etc.). The provider already supplies the complete text at item/completed; publishing that snapshot losslessly lets downstream folding replace any truncated preview with the full message.

Impact

Every completed OpenAI assistant item now converges to a complete persisted log entry. Messages that never receive a provider completion event can still only expose the streamed prefix.

Validation

• bun test tests/openai-app-server-event-bridge-*.test.ts tests/openai-app-server-provider-fixtures.test.ts (60 pass)
• bun run tc
• bun run build
• changed-file format check and source lint
• API fold regression suite in the parent repo (16 pass)