feat: check quality and target in the regression gate

[nucliweb]

Summary

Closes the three regression-gate items on the ROADMAP.

Check	Before	After
Size	bytes > +2% fails	unchanged
Quality	not checked	ssimulacra2 drop > 0.5 points fails as QUALITY, even if the file is smaller
Target	not checked	a run target different from the baseline's fails with an explicit error, comparing nothing
CSV	split on commas	writer quotes fields (RFC 4180), parser honours quotes

• baseline.json now stores { bytes, ssimulacra2 } per image and codec. It is regenerated from the latest main gate run's results.csv (amd64, like CI); bytes change only in the AVIF and HEIC pairs the trixie move had already shifted (≤0.41%).
• regression.yml defines TARGET once for the comparison and the check; the comment reads "no size or quality regression".
• check-regression.js gains --quality-tolerance, applies --target in check mode, and prints quality next to size.
• regression.js on main held a raw NUL byte as its key separator, so git treated the file as binary and never showed its diffs. It is now the \u0000 escape: same behaviour, plain text.
• Docs: both READMEs describe the three checks, and the baseline regeneration instructions use the CI CSV, since sizes differ slightly across architectures.

Test plan

• TDD: 10 tests in regression.test.js (7 failed before the change); node --test 51 pass, 0 skipped
• A CSV with JPEG XL at −0.8 points and −2.2% bytes is reported as QUALITY, exit code 1
• Checking at target 85 against the target-90 baseline fails with an explicit message, exit code 1
• A real compare-codecs run on an image named sunset, beach.png writes a quoted field that parses back correctly
• The workflow's commands inside the Docker image (arm64) pass; cross-architecture quality differences are ±0.04 points, well inside the 0.5 tolerance
• actionlint passes
• Gate on this PR (amd64) passes against the regenerated baseline

[github-actions]

Codec comparison ✅ no size or quality regression

📥 Download the codec report (unzip and open report.html locally). Available until 2026-12-31, when GitHub deletes the artifact.

Size check

Size regression check (target 90, size tolerance 2.0%, quality tolerance 0.5 ssimulacra2)

  ok         kodim04.png  JPEG XL  base 120534 → 120534  (+0.00%)  ss2 90.23 → 90.23  (+0.00)
  ok         kodim04.png  jpegli  base 130656 → 130656  (+0.00%)  ss2 90.02 → 90.02  (+0.00)
  ok         kodim04.png  HEIC  base 141631 → 141631  (+0.00%)  ss2 89.77 → 89.77  (+0.00)
  ok         kodim04.png  AVIF  base 150019 → 150019  (+0.00%)  ss2 90.11 → 90.11  (+0.00)
  ok         kodim04.png  WebP  base 173748 → 173748  (+0.00%)  ss2 89.81 → 89.81  (+0.00)
  ok         kodim04.png  mozjpeg  base 201372 → 201372  (+0.00%)  ss2 90.13 → 90.13  (+0.00)
  ok         kodim19.png  JPEG XL  base 136558 → 136558  (+0.00%)  ss2 90.02 → 90.02  (+0.00)
  ok         kodim19.png  jpegli  base 147032 → 147032  (+0.00%)  ss2 90.02 → 90.02  (+0.00)
  ok         kodim19.png  AVIF  base 148498 → 148498  (+0.00%)  ss2 89.74 → 89.74  (+0.00)
  ok         kodim19.png  HEIC  base 155322 → 155322  (+0.00%)  ss2 89.94 → 89.94  (+0.00)
  ok         kodim19.png  WebP  base 173918 → 173918  (+0.00%)  ss2 89.96 → 89.96  (+0.00)
  ok         kodim19.png  mozjpeg  base 213006 → 213006  (+0.00%)  ss2 89.92 → 89.92  (+0.00)
  ok         kodim23.png  JPEG XL  base 74113 → 74113  (+0.00%)  ss2 89.59 → 89.59  (+0.00)
  ok         kodim23.png  jpegli  base 97039 → 97039  (+0.00%)  ss2 90.25 → 90.25  (+0.00)
  ok         kodim23.png  AVIF  base 111556 → 111556  (+0.00%)  ss2 89.74 → 89.74  (+0.00)
  ok         kodim23.png  mozjpeg  base 130349 → 130349  (+0.00%)  ss2 89.85 → 89.85  (+0.00)
  ok         kodim23.png  WebP  base 140374 → 140374  (+0.00%)  ss2 89.04 → 89.04  (+0.00)
  ok         kodim23.png  HEIC  base 157442 → 157442  (+0.00%)  ss2 90.22 → 90.22  (+0.00)

OK: no size or quality regressions.