Add an end-to-end generation test using a tiny randomly initialized decoder model so CI exercises the public randkv.transformers.generate call without downloading a checkpoint.
Acceptance criteria:
- Construct the model and tokenizer inputs locally; no network access.
- Generate enough tokens to trigger at least two eviction rounds.
- Assert the configured cache bound and monotonic absolute position tracking.
- Run on CPU in the existing Python 3.10/3.13 matrix.
- Keep the test deterministic and under 10 seconds per CI job.
Add an end-to-end generation test using a tiny randomly initialized decoder model so CI exercises the public
randkv.transformers.generatecall without downloading a checkpoint.Acceptance criteria: