V2: overlapping reasoning fragments from OpenAI Responses adapter kill the execution as unknown #36892

Summary

A live production turn failed with:

session.execution.failed
error: { type: "unknown", message: "reasoning start before end: rs_0b8a8fd5e8c6a27e016a564fcbe6b481918eb0f7408aa4d6d7:0" }

The LLM event publisher enforces at most one open reasoning fragment (fragments("reasoning", ..., single = true) in packages/core/src/session/runner/publish-llm-event.ts). The OpenAI Responses adapter emitted a reasoning-start while a prior reasoning fragment was still open, and the publisher terminated the entire execution with an untyped defect.

Evidence

Session ses_09f05df32ffewmX7otmzvMPbKb (next channel), 2026-07-14:

15:03:40  session.step.started        openai/gpt-5.6-sol, variant high
15:03:40-15:03:52  two reasoning parts durably started and ended
15:03:52  session.execution.failed    unknown: reasoning start before end: rs_...:0

The turn had very high context pressure (~365k cached input tokens); auto-compaction ran on a later turn. The raw provider stream was not retained, so whether OpenAI emitted overlapping reasoning summary parts or the adapter derived overlapping normalized fragment ids is not determinable after the fact (adjacent to the diagnostics gap in #36766).

Where

• packages/llm/src/protocols/openai-responses.ts keys reasoning fragments as ${itemID}:${summaryIndex}, opening :0 at output_item.added and additional fragments per summary part. Some interleaving (for example a second reasoning item added before the first item's open fragment concludes) can produce reasoning-start while another fragment is open.
• packages/core/src/session/runner/publish-llm-event.ts responds to that with Effect.die, so one malformed fragment boundary fails the whole execution as unknown instead of a typed provider error.

Suggested Direction

1. Adapter invariant: the OpenAI Responses lifecycle should always emit reasoning-end for the open fragment before starting the next one. Audit multi-item and multi-summary-part transitions.
2. Publisher degradation: on an overlapping reasoning-start, close the open fragment (preserving its accumulated text) and continue, or fail with a typed provider.invalid-output error instead of dying as unknown. A reasoning boundary defect should not be indistinguishable from an internal crash.
3. Regression test: publisher receiving reasoning-start(a), reasoning-start(b) without reasoning-end(a) must not produce an untyped execution failure.