[wenh06]

Summary

Distill the parallel-execution design document from the maintainer's scratch area into a finalized documentation page (docs/source/parallel_execution_design.rst) — no intermediate revision or review history, describing the shipped system:

• requirements and the three schools of prior work;
• the two state-ownership invariants (I1 client-task enclosure, I2 server single-threadedness) and the arithmetic-equivalence corollary;
• the executor abstraction (in-order map, main-thread completion callbacks, fail-fast, context-manager resource handling) and why map-instead-of-futures;
• the per-round task decomposition, including the independence of uplink skipping from downlink broadcast and from training/evaluation;
• the backend table with measured speedups;
• the reproducibility ladder (L0 / L1 per-client generators / L2 rejected) and the configuration surface;
• implementation notes (named model_state exchange, the APFL fix found by the bitwise migration harness, the L1 loader-rebuild choice, the legacy-hook shim) and the outlook for the processes and ray backends.

Wired into the Advanced Topics toctree and cross-linked from parallel_execution.rst. zh_CN translation included.

Note: this PR is based on master before #68; the two zh build warnings on the theory pages are fixed there and are unrelated to this change.

Validation

• English and zh_CN docs builds: zero warnings from this page (the two zh warnings on the theory pages are the pre-docs(zh_CN): fix the three inconsistent-reference warnings in the zh build #68 ones).
• The design page renders in both languages; cross-links resolve.