[Mikyx-1]

Checklist

• I have reviewed and understand the changes in this PR.
• Confirmed that cargo run-checks has been executed.
• Made sure the book is up to date with changes in this PR.
• Added or updated changelog entries for user-visible changes, if applicable, following the
  changelog policy in CONTRIBUTING.md.

Related Issues/PRs

Closes #6223.

Changes

Use U8 confusion codes when supported, with the original integer expression as the fallback.
Keep count reductions and metric behavior unchanged. Add regression tests and a changelog entry.

The four codes need only two bits. U8 works with existing tensor operations; packing four codes
per byte would require separate encoding/counting logic and performance measurements.

Benchmarks

Complete Precision updates, threshold 0.5, micro reduction. Flex CPU, Intel i5-12400F,
release mode, Rust 1.98.0, default I32, one worker. Medians of nine paired rounds
(fifteen for 262,144 classes):

Batch × classes	Previous µs	U8 µs	Speedup
32 × 10	5.29	5.05	1.05×
32 × 1,000	37.70	21.36	1.77×
256 × 4,096	1,441.43	766.78	1.88×
1,024 × 4,096	18,825.03	7,589.64	2.48×
1 × 262,144	333.44	170.97	1.95×

At 256 × 4,096, additional allocation peak falls from 9 MiB to 6 MiB. At 1 × 262,144,
macro gains are only 1.02–1.11× across the three metrics; top-k performance is largely unchanged.

Testing

• cargo test -p burn-train --lib metric:: --no-default-features --locked: 281 passed.
• cargo run-checks: passed with Flex; existing ignored tests remained skipped.
• Paired release benchmarks: 822 configurations passed exact-output checks. Allocation counting
  ran separately from latency measurements.
• GPU execution and the unsupported-U8 fallback branch were not tested.

AI Usage

Codex autonomously implemented and tested the change, ran validation and benchmarks, and updated
the changelog following user instructions. It drafted the issue and PR descriptions and created
the issue and this PR.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 66.15%. Comparing base (dd31726) to head (54274d8).
⚠️ Report is 1 commits behind head on main.

❌ Your project check has failed because the head coverage (66.15%) is below the target coverage (80.00%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6224      +/-   ##
==========================================
- Coverage   66.16%   66.15%   -0.02%     
==========================================
  Files         878      878              
  Lines      127748   127748              
==========================================
- Hits        84528    84513      -15     
- Misses      43220    43235      +15     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.