[renanmpimentel]

Summary

Add a test that Phoenix.Token.sign/4 forwards :key_length to the key generator. It compares deterministic tokens made with lengths 16 and 32, verifies the 16-byte token with the matching option, and checks that the mismatched length is rejected.

Why

The documentation lists :key_length as a signing option, but the existing token tests compare only :key_iterations and :key_digest. The key defaults test explicitly uses the default length (32), so discarding :key_length in sign/4 leaves the entire existing suite green.

The regression was applied in an isolated worktree:

State	Result
Existing suite + sign/4 discards :key_length	1,111 passed
New test + same regression	Assertion fails: tokens for lengths 16 and 32 become identical
New test + correct production	Passes

The permanent diff changes only test/phoenix/token_test.exs.

Verification

• mix format --check-formatted test/phoenix/token_test.exs: passed
• mix compile --warnings-as-errors: passed
• mix test --warnings-as-errors: 1,112 passed, 33 excluded
• git diff --check: passed

One earlier full run had a timing failure in the unrelated crashed channel propagates exit test (200 ms assertion window); that test passed in isolation and the full suite passed on rerun.

Found with Supertest, an audit method for test effectiveness.