Add optional regression tests for 4 confirmed bug behaviors

[artuskg]

Summary

This PR adds a small, MLX-backed optional regression test suite for four high-impact behavior bugs we observed in voxmlx.

The tests are intentionally synthetic and fast, and they are gated behind VOXMLX_ENABLE_MLX_RUNTIME_TESTS=1 so default CI remains lightweight.

Why this PR

We wanted a precise way to validate whether the following issues are real bugs (not just quality drift or perf noise), and to prevent regressions after fixes.

The 4 bugs (with test mapping)

1) RotatingKVCache concat path can exceed max_size for multi-token appends

• Behavior: when appending S > 1 tokens with concat updates, cache length can exceed its cap and keep growing above the intended window.
• Why it is a bug: windowed KV cache semantics are violated; this affects memory and can change attention context unexpectedly.
• Tests:
  • test_rotating_kv_cache_concat_respects_max_size
  • test_rotating_kv_cache_concat_keeps_expected_tail_for_multi_token_append

2) encode_step() cache window should come from encoder config (sliding_window), not a large hardcoded constant

• Behavior: cache is created with an oversized fixed window (historically 100000) instead of model-config window.
• Why it is a bug: ignores model/runtime contract and can cause unnecessary memory/computation growth for long streams.
• Test:
  • test_encode_step_uses_encoder_sliding_window_for_cache_size

3) Offline encode() trims from the head instead of the tail

• Behavior: when frame count is odd or not divisible by downsample factor, leading frames are dropped.
• Why it is a bug: shifts alignment and makes offline behavior inconsistent with incremental path expectations.
• Test:
  • test_encode_trims_trailing_frames_not_leading_frames

4) Offline generate() drops the final pending token when loop ends without EOS

• Behavior: decode loop uses a pending-token pattern but does not flush the last pending token at natural loop termination.
• Why it is a bug: deterministic one-token truncation at end-of-audio in non-EOS termination path.
• Test:
  • test_generate_flushes_final_pending_token

Test file

• tests/test_mlx_runtime_optional.py

How to run

VOXMLX_ENABLE_MLX_RUNTIME_TESTS=1 python3 -m unittest -v tests.test_mlx_runtime_optional

Notes

• These are regression tests only; they do not enforce any specific implementation strategy.
• They are designed to be stable and minimal, using synthetic tensors and stubs where possible.

[artuskg]

Superseded by four focused fix+test PRs: #3 (KV cache trim), #4 (encoder sliding window), #5 (offline encode trim direction), and #6 (generate final-token flush). Closing this broad test-only PR to keep review scope tight.