fix(opencode): make text/reasoning delta accumulation O(N) instead of O(N²)

[hardes11]

Issue for this PR

Fixes#30067

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Resubmission of #30058 by meitalbensinai (auto-closed for missing PR template; commit cherry-picked unchanged, authorship preserved).

Replaces text += delta accumulation in the reasoning-delta and text-delta handlers of session/processor.ts with a lazy chunked-text getter. Once anything reads .text between writes (UI render loop, NDJSON serializer, bus event broadcaster), the runtime flattens the rope string and the next += re-copies the whole buffer — O(N²) in cumulative part length. On thinking-mode models emitting 1500+ small reasoning tokens per turn, this grows per-step latency from ~6s to 100s+ by turn 50-80 (perf evidence in #30067: memmove at ~30% of CPU, GC pinned). Same shape as the fix that landed in vercel/ai for processUIMessageStream.

The getter is non-enumerable, so it does not leak through JSON.stringify/structuredClone; .text reassignment still works via the setter.

How did you verify your code works?

• Cherry-picked the original commit onto current dev and resolved the conflict in the import region (dev had since dropped an unused log variable; the incoming function is the only addition).
• Diff is one file, +48/-2; confirmed no += accumulation remains in either handler and no orphaned references were introduced.
• Ran a local harness (Bun 1.3.14) executing the exact installChunkedText implementation from this diff:
  • Functional: after 80k appended deltas, .text is byte-identical to naive += accumulation; the getter caches the joined string after first read; the setter resets the buffer and subsequent pushes accumulate correctly; _chunks is absent from JSON.stringify and object spread (non-enumerable, no leakage).
  • Performance: with flatten-forcing reads (JSON.stringify every 100 deltas), the chunked path is ~2.9x faster at 80k deltas / 1.6MB accumulated text, and the naive path re-copies the whole buffer on every append once a read forces flatten, so the gap widens with length.
• Did not build the app or run the full test suite; the harness exercises the changed mechanism in isolation, not the whole binary.

Screenshots / recordings

n/a

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍