Stream failures between retry attempts leave reasoning/text parts permanently unclosed in transcripts

While looking into sessions that hit repeated provider 503s, I noticed stored part records that have time.start but never get a time.end.

Cause as far as I can tell: in packages/opencode/src/session/processor.ts, when an attempt fails, the retry loop resets ctx.currentText = undefined / ctx.reasoningMap = {} without finalizing the parts already persisted during streaming. The sealing logic lives in cleanup(), which only runs after the whole retry chain settles - so every failed attempt in the middle of a chain leaves dangling blocks behind.

On one long-running install I count ~970 such parts (reasoning and text). No accounting damage (usage comes from step-finish), but transcripts accumulate blocks that never officially ended, which is visible in the session view after unstable-provider stretches.

Happy to send a PR if this seems worth fixing.