[novkien]

Overview

Adds a reproducible Flash Next streaming benchmark harness for explicit proxy endpoints. It prepares exact prompt lengths, checks native counters and timings, preserves raw SSE evidence, and records artifact/runtime provenance so incomplete or changed campaigns are not reported as acceptance results.

Addresses #9

Validation

• python3 -m unittest -v tests.test_bench_flash_next (13 tests passed)
• python3 -m py_compile scripts/bench_flash_next.py
• git diff --check
• Existing live screening is recorded on the issue; this tooling PR makes no optimization speedup 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]

CI update for ae14eb10fc26138f54e7c66b473ad15ace533598

The fork's Lint check failed. Flake8 reports NP100 for print() at scripts/bench_flash_next.py:1475 and E402 for the import after sys.path.insert at tests/test_bench_flash_next.py:10.

Labeler and model-naming passed. Editorconfig and two Python type-check jobs were still pending at this observation. The focused 13-test suite, Python compilation, and git diff --check passed locally. This reviewed head remains open and unmerged while the two lint findings are corrected and reviewed.

[novkien]

Validation update for 1ebe411a30851671e59d616fa20d16a66a3bb670

The two Lint findings are corrected on this head. The PR Lint workflow now passes (run 37094369075). Editorconfig, labeler, and model-naming also pass. The two Python type-check jobs are still pending; scoped local type-check, the 13 focused tests, Python compilation, and git diff --check passed.

The PR remains measurement support only and makes no performance improvement claim.

[novkien]

Follow-up finding: generated fixture padding

A live native preparation reproduced a failure for a generated prose fixture at 2,048 target tokens with seed 731; the nearest lower count was 2,035. In scripts/bench_flash_next.py:942, fit_prompt() derives the accumulated padding with content[len(base_content):]. Padding is inserted before the final question, so base_content is not a prefix of the candidate; when another padding iteration is needed, the extracted suffix is from the question instead of the padding. The generator then fails to reach the exact count.

This is a real limitation in the merged runner's generated-fixture path. PR #13 is already merged; the assigned follow-up is to correct suffix tracking and add a multi-step regression/preparation check under issue #9. This does not change the PR's status as measurement tooling and does not establish a speedup.