[mremond]

Why

rand:uniform/1 could not be seeded and no other rand function existed, so seeded programs such as PropEr runs drew different numbers on every run.

What changed

• Programs can seed rand and get the same numbers, states and errors as OTP 29 for the same seed and algorithm.
• The rest of OTP's rand module works: uniform/0, uniform_s, uniform_real, bytes, normal, shuffle, jump and exported or explicit states, for every algorithm.
• math:exp/1, math:log/1 and math:pow/2 are available, with OTP's errors.
• trunc, round, ceil and floor return the exact integer for floats beyond 2^63 instead of a clamped one.
• Cowlib's seeded date property now runs natively with the same traces as on BEAM.

Before and after

Before: A program calling rand:seed/2 failed to build because the rand module was not available, and two runs of rand:uniform(100) printed different numbers.

After: A seeded program prints the same numbers on every run, equal to OTP 29.1.2's output, and Cowlib's date property passes 6,000 trials per engine with identical traces.

This changes what compiled programs compute and print in a terminal; nothing on screen changed, so the evidence is program output and test results.

Checked

• A seeded demo program built natively, run twice and compared with OTP 29.1.2: Both runs print the same numbers, equal to OTP's output line for line
• rand_regression: every algorithm and seed form against values generated on OTP 29: Passes 10 of 10 runs natively and 10 of 10 on BEAM, with the math and float fixtures
• Cowlib 2.22.0 prop_parse_date with native PropEr, three seeds run twice: 6,000 of 6,000 trials per engine; native and BEAM traces identical
• Full local gate: clippy, Rust tests, allocator check, end-to-end suite: 2,142 Rust tests and 471 of 473 end-to-end tests pass; two Onecached tests skip as before
• Cost on a trivial program and on draws: About 0.1 s more build time and 0.57 MB more executable; uniform(100) 145 ns instead of 24 ns

How it was checked

cargo build --release --locked, then cargo clippy --workspace --all-targets --locked -- -D warnings

cargo test --workspace --locked --no-fail-fast (2,142 passed)

python3 scripts/test-allocator-config.py

./target/release/nera test -j 1 --timeout 30 (471 passed, 2 skipped, 0 failed)

nera test rand_regression math_exp_log_pow float_integer_large rand_uniform, ten times in a row, and the same four modules on OTP 29 ten times

python3 tests/integration/cowlib/native_date.py --pin 2.22.0 with the pinned PropEr and OTP 27.3 rand and ordsets sources

After bringing the branch up to date with the atom-quoting fix on main: rebuild, clippy, the runtime tests (755 passed) and the rand, math, float, sqrt and atom-quoting fixtures pass

Risk

Risk above low: Blast radius, medium — Every program compiles the rand module; four float conversions and three math functions change; Behavior, medium — New rand functions are opt-in; uniform/1 now uses OTP's generator and is slower; Novelty, medium — A 2,188-line module ported from OTP beside the existing stdlib

Full risk assessment

Dimension	Level	Why
Blast radius	medium	Every program compiles the rand module; four float conversions and three math functions change
Behavior	medium	New rand functions are opt-in; uniform/1 now uses OTP's generator and is slower
Undo	low	Revert and done; no saved data or settings move
Evidence	low	Seeded demo and Cowlib's date property shown matching OTP end to end
Sensitivity	low	No sign-ins, keys, saved data or releases
Novelty	medium	A 2,188-line module ported from OTP beside the existing stdlib
Dependencies	low	No new or updated packages

Not done

A normal draw takes about 2 µs natively (37 ns on BEAM): each draw rebuilds a 256-entry table because Nera has no constant literal storage.

Every program compiles rand, about 0.1 s and 0.57 MB more, even when it never calls it.