[OskarEichler]

Summary

Coerce byte lengths once before passing them to the selected backend, matching the integer conversion accepted by Random.bytes and avoiding false partial-read errors.

Reproduction and verification

SecureRandom.bytes(3.5).bytesize # 3
size = Object.new
def size.to_int; 3; end
SecureRandom.bytes(size).bytesize # 3

The baseline raises NotImplementedError even though the native backend returned the requested coerced number of bytes. Integers retain a fast path; random_bytes/hex and other formatter methods still call gen_random directly.

• Existing rbenv exec rake test: 3 tests / 68 assertions, zero failures/errors on the baseline, this isolated branch and the cumulative release-based consumer, Ruby 4.0.6/macOS. The existing disabled PID-recycling stress test was not enabled.
• Combined candidate: 1,539 length/formatter checks, 3 fallback checks and 13 error/coercion/Ractor checks pass. Package build/extraction and existing rdoc -C checks pass (100% of its documentation inventory).
• Paired benchmark, 10,000 16-byte calls per round: bytes median 0.010524s baseline / 0.010591s candidate; random_bytes 0.010651s / 0.010635s. Both allocate 10,000 objects. These small timings do not support a speedup claim.
• No tests/specs added or modified under the consumer repository's explicit policy. External scratch reproductions were used.

Breaking changes and limitations

Previously rejected fractional/to_int lengths now work; non-convertible inputs remain TypeError, with a clearer size diagnostic. Invalid Float conversion now agrees with Random.bytes. No RNG backend, random-byte content, or entropy algorithm changes.

No production execution, statistical randomness claim, or cross-platform CI claim. Older Ruby/native fallback environments need upstream CI. The shared Random::Formatter implementation belongs to Ruby itself and is not changed here.