Docs: record the drain-queue QoS inheritance gap and the measured cost of #70's round-trip

[mansbernhardt]

Summary

Doc-only. Adds Update 28 to Docs/test-determinism-executor-drain.md recording the outcome of #70 (closed):

• The drain queue inherits the submitting thread's QoS, so a task resumed from a background thread runs its drive job at QOS_CLASS_BACKGROUND — a real gap in the non-starvable contract, measured directly on that branch. Victimless today: the wedge that motivated it was the 1.0.18 deadlock.
• The floor alone is unsafe (premature settle on the 3-core runner — yielding children hop through the global executor and are invisible to outstanding); the branch's Task.yield() round-trip fixes that but costs, measured on 1.0.20 with 12 interleaved paired full-suite --parallel runs: wall time unchanged, summed per-test in-flight time +14 % median / +20 % mean, slower in 10/12 pairs, heavy tail, spread uniformly across every test (one hop per settle()).
• What to do if a victim ever appears.

🤖 Generated with Claude Code