fix(ai): support streams without finish reasons - #40965

Summary

• honor compatibility.requireFinishReason, preserving strict behavior when unset or true
• synthesize an unknown terminal finish after non-empty clean EOF when the option is false
• make protocol halt finalization optionally effectful so OpenAI Chat can finalize accumulated tool inputs and preserve usage
• keep empty streams, malformed events, incomplete tool identity, provider errors, and transport failures as errors
• verify Core continues the agentic loop from recorded tool calls even when the finish reason is unknown

Behavior

For requireFinishReason: false, a non-empty stream that ends without a terminal reason completes with step-finish and finish events whose normalized reason is unknown. OpenAI Chat finalizes valid and malformed local tool inputs before emitting that finish. Core continuation remains content-driven, so valid local tools execute and cause the next model step without requiring a synthesized tool-calls reason.

Testing

• packages/ai: bun typecheck
• packages/core: bun typecheck
• packages/core: bun test test/session-runner.test.ts (144 pass)
• targeted generic and OpenAI Chat compatibility tests (10 pass)
• Gemini and Bedrock protocol regressions (64 pass)
• full packages/ai suite: 481 pass, 28 skipped, 4 unrelated base failures
• Prettier and git diff --check

The existing full AI suite failures are Azure URL expectations with a duplicated /v1 and stale OpenRouter recorded request bodies expecting no usage option; neither path is changed by this PR.