Goal

Make local cancellation boundaries explicit in nested LLM stream diagnostics so #721-class local aborts can be distinguished from unknown iterator failures when abort provenance exists.

Scope

In scope:

• Add a stream diagnostic hook such as recordAbortState / abort provenance capture.
• Merge existing assistant abort diagnostics or interrupt metadata into llm_trace.stream.abort.
• Classify local_abort as high confidence only when the abort signal is aborted and provenance is present.
• Preserve watchdog precedence when watchdog and abort evidence overlap.

Out of scope:

• Changing retry behavior.
• Changing user-facing abort copy.
• Treating a bare aborted signal as local abort without provenance.

Relevant files or context

Related work:

• [Bug] First-turn tool execution aborted without recorded source — model stops, user must resend #721
• [Task] Add LLM stream failure diagnostics v2 to session exports #760
• feat(session): add llm stream failure diagnostics #771

Likely files:

• packages/opencode/src/session/llm.ts
• packages/opencode/src/session/processor.ts
• packages/opencode/src/session/llm-trace/types.ts
• packages/opencode/src/session/llm-trace/recorder.ts
• packages/opencode/src/session/llm-trace/stream-diagnostics.ts
• packages/opencode/test/session/llm-trace.test.ts
• packages/opencode/test/session/llm.test.ts

Current gap verified during PR #771 review: recordIteratorError passes abortProvenancePresent: false, while the classifier already supports local_abort when provenance is present.

Verification

• Add local abort + iterator error coverage: provenance present => boundary: \"local_abort\", confidence: \"high\".
• Add watchdog/abort overlap coverage: watchdog evidence still wins.
• Add missing provenance coverage: aborted signal + iterator error + no provenance remains unknown / low.
• Run targeted session trace tests and typecheck.

Execution mode

Investigate and propose a plan first — the agent must post the plan as an issue comment and wait for an explicit "approved" comment before writing code or opening a PR.