test: cover Bits1_5 in TurboQuant test matrices - #10448

Summary

Nine bit-width matrices in lib/quantization/src/turboquant/quantization.rs omit TQBits::Bits1_5, so that variant is never exercised by them.

This is the same enumeration gap #10390 closed for quantize_output_byte_length: the quantized_size double-padding bug was invisible because the one test that would have caught it did not list the affected variant. The rest of the module still has that gap, while score_precomputed_batch_matches_single already enumerates Bits1_5 — so the inconsistency sits inside a single file.

Test-only change; no production code is touched.

Changes

Bits1_5 added to the seven matrices where it applies:

• quantize_extreme_values
• quantize_deterministic
• quantize_zero_vector
• quantize_non_power_of_two_dims
• pack_unpack_vector_roundtrip
• pack_unpack_vector_uniform_indices
• score_precomputed_dispatches_all_bit_widths — the name already claims to cover all bit widths

The two Unpadded rotation tests cannot take it, and now say so:

• unpadded_rotation_keeps_padding_zero and unpadded_rotation_matches_padded_for_padding_free_dims both build quantizers with TQRotation::Unpadded, and TurboQuantizer::new asserts Bits1_5 requires TQRotation::Padded — Bits1_5 rotates into its x1.5 padding, so an unpadded rotation would leave half the codes carrying nothing. Adding it there panics; each list now carries a one-line comment naming the invariant, so the omission reads as deliberate.
• unpadded_rotation_matches_padded_for_padding_free_dims also claimed multiples of 8 are padding-free "for every supported bit width". That stopped being true when Bits1_5 was added — padded_dim(8) is 16 — so the comment now names the bit widths it actually covers.

Left alone as out of scope: higher_bits_reduce_error asserts the ordering MAE(Bits4) <= MAE(Bits2) <= MAE(Bits1). Placing Bits1_5 in that chain is an accuracy claim that deserves its own justification rather than being folded into a coverage change.

Verification

Run on stable 1.98.0, the toolchain .github/workflows/rust.yml uses.

Before the change the same filter reports 89 tests; the extra one is the new score_precomputed_dispatches_all_bit_widths::bits1_5 case.

AI contribution

Prepared with Claude Code: it audited the bit-width matrices in the module, made the edits, and ran the commands above.

The first attempt added Bits1_5 to unpadded_rotation_keeps_padding_zero as well; the test run caught the Bits1_5 requires TQRotation::Padded assertion in TurboQuantizer::new, which is what established that both Unpadded tests must keep their current lists. The scope call on higher_bits_reduce_error was made against TQBits::bit_size and the centroid tables.

All Submissions:

• My PR targets the dev branch (not master) and my branch was created from dev.
• Have you followed the guidelines in our Contributing document?
• Have you checked to ensure there aren't other open Pull Requests for the same update/change?

Changes to Core Features:

• Have you added an explanation of what your changes do and why you'd like us to include them?
• Have you written new tests for your core changes, as applicable?
• Have you successfully ran tests with your changes locally?