V2: define typed LLM errors for consumer-owned retry policy

[rekram1-node]

Context

The LLM package should not decide retry policy, but opencode needs useful typed failures to implement its own policy. AI SDK errors are currently hard to interpret, especially for midstream provider failures.

Scope

• Review packages/llm/src/schema/errors.ts and executor error mapping.
• Prioritize context overflow, provider/API errors, transport failures, auth/config failures, and midstream provider failures.
• Preserve raw provider details where useful without hardcoding brittle provider-specific status semantics.
• Remove retry-policy language from public errors if it implies package-owned retries.

Acceptance Criteria

• Consumers can distinguish context overflow from generic provider failure.
• Midstream provider failure is surfaced ergonomically.
• Errors include enough status/header/body/provider detail for opencode retry policy.
• No error API implies automatic retries inside @opencode-ai/llm.

[opencode-agent]

Audit: several recent PRs chip away at typed LLM/provider failures (#35671 Z.AI overflow classification, #36130 nested OpenAI stream errors, #36163 typed provider/compaction errors), but this still looks like a broader error-contract issue. Keeping open.