runtime: materialize channel resumes before P transfer

[cpunion]

Summary

• add a fixed-enum, frame-local ResumeCleanupPlan to the unified bounded runner
• retire direct-channel and multi-case select hchan waiters, exact source generations, claim and result lease before runnable publication
• resume from a compact P-neutral packet, including transfer to another P and prompt task cancellation after migration
• keep the plan POD-only: no callback, interface, Future/task object, function-address lookup, or permanent G field
• generalize the packet small-result union for poll and channel status while preserving its 52-byte align-4 cross-target layout

The runtime switch handles one logical select case per ExecutorRunStepMaterialize; common core confirmation and recycle are also one exact generation per reduction. The existing hchan buffer/closed-queue reconcile helpers can still process multiple peer waiters internally; splitting those loops into an explicit subcursor is the next strict-cost PR and is documented rather than overclaimed here.

Validation

• runtime core: race + shuffle
• production typed hchan adapter: native race + shuffle
• production typed hchan adapter: actual JS/WASM execution
• compiler coroutine analysis: race + shuffle
• isolated program driver and native fleet tests
• focused coroutine channel lowering tests

Based on the post-#50 llvm-coro baseline, which includes current xgo-dev/main.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.