[tamnd]

The "instrumented libraries that have to be right" job on main fails on zlib at -O2 with the short-circuit pass said it preserved the loop forest of gz_decomp and did not [E0652]. It reproduces on the Linux box with rucc -S -O2 -Zverify-each gzread.c, with or without -fsafety.

The forest was right. Short-circuit keeps the loop forest through each merge with Loops::merged, and a forest kept that way lists a loop's blocks in the order it found them before the merge. A fresh build after the merge found the new latch second rather than last. Printing both forests for gz_decomp showed that one block moving in the list and nothing else. The -Zverify-each check compared the two with ==, so it took the order for a difference. The debug assertion in clear_merging already compared them with Loops::summary, which sorts every list, and that is what the check now uses too. summary is no longer debug only.

The new test in analysis.rs is the smallest shape that does it, a loop of three blocks whose latch is folded into the block above it. It fails on the old check and passes now.

Checked on the Linux box: cargo test -p rucc-opt (1,938 unit tests and the integration tests) passes, cargo clippy -p rucc-opt --all-targets -- -D warnings and cargo fmt --all --check are clean, and every zlib 1.3.1 source compiles with -O2 -Zverify-each.