Report low-confidence Sinter decodes as discards

[danielgaskins]

Summary

• append Sinter's discard byte to compiled decoder results
• set the discard byte when Tesseract reports low confidence
• document errors + discards as the conservative logical failure count
• add direct-output and Sinter-sampler regression tests

This addresses the compiled-decoder portion of #297. The pinned Sinter file-decoder API reads exactly the observable bytes for each shot and has no decoder-controlled discard channel, so decode_via_files remains unchanged. The multipass decoder in #256 will need equivalent propagation when it is integrated.

Testing

• clang-format --dry-run --Werror src/tesseract_sinter_compat.pybind.h
• focused low-confidence tests: 2 passed
• bazel test --jobs=1 //src/...: 12 of 13 targets passed; the pre-existing unseeded test_sinter_decode_surface_code exceeded its statistical threshold locally

[danielgaskins]

Could someone with Actions access rerun the failed jobs? The macOS failure came from the existing unseeded test_sinter_decode_surface_code statistical threshold. The new deterministic discard tests passed, as did formatting and all wheel builds.

[mhucka]

Could someone with Actions access rerun the failed jobs? The macOS failure came from the existing unseeded test_sinter_decode_surface_code statistical threshold. The new deterministic discard tests passed, as did formatting and all wheel builds.

Done. FYI that I had to rerun it twice, actually; the first rerun saw macos succeed but linux fail; the second rerun saw the linux tests succeed. (Mentioning it in case this is useful for additional development.)

[danielgaskins]

Thanks for flagging this. The compatibility tests were still using Sinter's legacy sample_decode helper, which treats the new discard byte as observable data. I updated them to use Sinter's discard-aware sampler. The Bazel target now passes in 10 consecutive runs.