Make AMD64 field operations safe for in-place use

[JasonGross]

Summary

• snapshot input limbs before executing the optimized schedules in the 43 affected AMD64 implementations
• preserve support for both out1 == arg1 and out1 == arg2
• document the in-place calling contract
• add an exhaustive regression test covering all 620 shipped assembly implementations

The regression assembles each implementation, compares its distinct-buffer result with the corresponding generated C function, and then verifies both supported output/input aliasing forms. This addresses Scrutineer finding #2521.

Testing

• fiat-amd64/test_aliasing.py
• CC=clang fiat-amd64/test_aliasing.py
• python3 -m py_compile fiat-amd64/test_aliasing.py
• git diff --check

Both GCC and Clang runs report:

All 620 assembly files support output/input aliasing

The extracted Coq checker was not available locally because the isolated worktree lacked etc/coq-scripts/Makefile.vo_closure; clean CI should formally revalidate the modified hints. Existing performance ratio/footer metadata also predates the snapshot prologues and should be rebenchmarked before marking this ready.

Authorship note: this was researched and written by an AI coding agent
(OpenAI Codex), working on Jason Gross's behalf; Jason reviews what is
posted from this account.

[JasonGross]

Local review found no further patch changes necessary. I audited all 43 snapshot prologues for exact input coverage, stack-pointer balance, and separation between existing scratch space and the new snapshots. GCC and Clang runtime tests both pass all 620 assembly implementations. I also ran an alias-aware symbolic equivalence checker over all seven affected operation groups; every group passed the ordinary disjoint-buffer equivalence check and the out1 == arg1, out1 == arg2, and combined-alias configurations.

GitHub currently shows no completed failures. The full test-amd64 job is still queued; the completed platform jobs exercise only the lite AMD64 set, which does not include these 43 large-limb snapshots.

Authorship note: this was researched and written by an AI coding agent
(OpenAI Codex), working on Jason Gross's behalf; Jason reviews what is
posted from this account.