[novkien]

Overview

Fixes generated fixture preparation when more than one padding step is needed. Padding is inserted before the final question, so it must be tracked directly; slicing the rendered content after the base prompt length can take text from the question instead of the accumulated padding. The regression test covers the multi-step case.

Addresses #9

Validation

• python3 -m unittest -v tests.test_bench_flash_next (14 tests passed)
• Python 3.11 flake8 with flake8-no-print on the two changed files
• Python compilation and git diff --check
• Live native token preparation passed for all 18 combinations: three fixture families, two lanes, and target lengths 2,048, 8,192, and 32,768, seed 731.

This is a correctness fix for benchmark preparation and makes no inference performance claim.

Requirements

• Contributing guidelines reviewed for this submission.
• AI usage disclosure: YES - Codex implemented and reviewed this change under the owner-approved Flash Next plan.

[novkien]

Validation update for e1959f0abf370813d5fdc7ff60697fe5c3e8dcd9

The multi-step padding fix and regression are on the reviewed head. CI Lint passed (run 37096328003); editorconfig, labeler, model-naming, and one Python type-check job passed. The second type-check job was still pending at this observation. Local 14-test suite, Python compilation, scoped type-check, and git diff --check passed. Live native token preparation passed all 18 combinations: three families, two lanes, and 2,048/8,192/32,768 tokens at seed 731.

This corrects benchmark fixture generation only; it makes no inference speedup claim.