[mremond]

Why

Natively, erlang:max/2 and min/2 turned the integer of a mixed integer and float pair into a new float, so max(2, 1.5) returned 2.0 and max(1, 1.0) returned 1.0, where BEAM returns 2 and 1.

Rounding the integer first also made integers above 2^53 compare equal to smaller floats.

What changed

• max/2 and min/2 return one of their two arguments unchanged, the first when they compare equal, as on BEAM.
• Every pair goes through the term order, which already compares integers and floats exactly, so the separate numeric arms are gone.
• The new max_min_numbers fixture covers mixed pairs, signed zeros, values around 2^53, 2^60, 2^63 and beyond 64 bits, and direct, apply, fun and guard calls; it runs natively and on BEAM.
• The runtime unit test that expected max(3, 2.5) to be 3.0 now expects the integer itself; the changelog and the tests README describe the change.

Before and after

Before: Natively, max(2, 1.5) was 2.0, min(1, 1.0) was 1.0, and max(9007199254740993, 9007199254740992.0) was a float.

After: Natively, max(2, 1.5) is 2, min(1, 1.0) is 1, and max(9007199254740993, 9007199254740992.0) is the integer, as on OTP 29.

The change is in runtime results that a program prints or matches, not in anything drawn on screen.

Checked

• A probe of 37 mixed integer, float and atom pairs through max, min, apply and fun calls, printed with exact float bits, natively and on OTP 29.1.2: 18 lines differed from OTP before the change; the native output is byte-identical to OTP after it
• The max_min_numbers fixture with the compiler built from the base version: Fails at its first line with {badmatch,2.0}
• The max_min_numbers fixture after the change, ten runs natively and ten on OTP 29.1.2: 10 of 10 pass on each
• Timing of max/2 over 20 million calls, eight interleaved runs before and after: Integer pairs about 4.3 to 4.5 ns per call either way; differences sit within this busy machine's noise

How it was checked

cargo build --release --locked -p nera-runtime, then -p nera, then the whole workspace: passed

cargo clippy --workspace --all-targets --locked -- -D warnings: passed

cargo test --workspace --locked --no-fail-fast: 2,139 passed, 0 failed, 1 ignored

python3 scripts/test-allocator-config.py: passed

nera test -j 1 --timeout 30: 467 passed, 0 failed, 2 skipped (external OneCached modules), max_min_numbers included

cargo test -p nera-runtime --lib bif::tests::bif_m ten times: 10 of 10, five tests each

nera test max_min_numbers ten times: 10 of 10; the same fixture under erl on OTP 29.1.2 ten times: 10 of 10

Checked on macOS arm64 only

Risk

Risk: low in every dimension.

Full risk assessment

Dimension	Level	Why
Blast radius	low	One runtime helper behind erlang:max/2 and min/2
Behavior	low	Corrects a result that differed from BEAM
Undo	low	Revert the commit
Evidence	low	Native output matches OTP 29 on 37 pairs; fixture runs natively and on BEAM
Sensitivity	low	No settings, data or sign-ins
Novelty	low	Reuses the existing exact term comparison
Dependencies	low	No dependencies added or updated