[prakashUXtech]

Last part of IC-2, after #590, #592 and #594. It adds the rest of the JPEG checks, each disabled until the task that implements it (lead.md rule 13). No Sources file changes.

What

JPEGFixtures gains two functions:

• progressive(f): a SOF2 stream (T.81 Annex G) carrying the same coefficients as baseline(f). The scans are:

  • DC first at Al 1, then the DC refinement, both interleaved.
  • Chroma AC by spectral selection, bands 1–5 and 6–63.
  • Luma AC 1–63 at Al 1, then refined to Al 0.

  Each AC scan holds one component over that component's own block grid (A.2.2).

• refineAC: the G.1.2.3 refinement coding.

  • A newly nonzero coefficient is coded as (R, 1) plus a sign bit. R counts only zero-history zeros.
  • A coefficient with history sends a correction bit, held back until the next symbol.
  • A ZRL goes out at the next nonzero once 16 zeros have passed, and never after the last new coefficient.

JPEGFuzzTests (IC-6):

• thousandMutationsThrowOrDecode: 1000 mutated reads and 200 mutated info calls.
• scanMutations: damages only the entropy-coded bytes, behind valid headers.
• seededValidStreamsDecode: 40 seeded random sizes, samplings and restart intervals, each within tolerance.

The seeds cover grey, 4:2:0 and 4:2:2, restarts, and EXIF in both byte orders. Each mutated case must throw a CodecError or decode.

JPEGReadTests, progressive half (IC-7):

• progressiveMatchesBaseline: grey and the three samplings. The progressive coding must read within 1/255 of the baseline coding and within the reference tolerance, and info must report progressive.
• acRefinement: AC values large enough that refinement carries every low bit and every ±1. It covers mutation .acRefinement. Zigzag positions 2–17 are 16 zero-history zeros in front of a coefficient with history, so the stream needs a ZRL with a held correction bit.
• damagedThrows now wraps dropFirst() in Data(...) so the slice starts at index 0.

JPEGWriteTests (IC-8):

Test	What it checks
qualityNinetyReopens	seeded gradients reopen within a mean of 3/255; alpha is 1
qualityRange	qualities 1 and 100 are accepted; 0, 101, -1 and Int.max throw unsupported; a side over maxSide throws tooLarge
startsWithSOIAndJFIF	SOI, then APP0 "JFIF\0", then EOI. SOF0 only, 8-bit, Y at 2x2 and Cb/Cr at 1x1. info agrees
annexKHuffmanTables	the DHT tables equal K.3 to K.6
qualityScale	covers mutation .qualityScale. Below 50 the scale is 5000/q, otherwise 200 − 2q, and each entry is clamp((K·scale + 50)/100, 1, 255). Quality 50 gives K.1 exactly and 100 gives all ones
lumaTables	covers mutation .lumaTables. The Y component's Tq names scaled K.1, and Cb/Cr name scaled K.2

Spec sources and readings

Sources are ITU-T T.81 (A.2.2, A.4, Annex G (G.1.1.1.1, G.1.2.1 to G.1.2.3), B.2.4.1, B.2.4.2, Annex K) and JFIF 1.02.

Places where the spec leaves a choice, and the reading I picked:

• AC point transform. The AC point transform divides toward zero. That is what makes the refinement's correction bit equal bit Al of |coefficient|.
• DC point transform. The DC point transform is an arithmetic shift right, and its refinement sends bit Al of the two's-complement value (G.1.2.1).
• Where correction bits go around a ZRL. In refinement scans, each ZRL is followed only by the correction bits of history coefficients before the 16th zero it skips, which is how the G.1.2.3 decoder reads them. My encoder sends a ZRL at any nonzero once 16 zeros have passed, so the bits it holds back always fall inside that span.
• Quality scaling. The packet asks for "the common 50-centred formula" without spelling it out. I used the one above.
• Write errors. For write, a quality outside 1…100 is unsupported, as the IC-1 doc comment says.

Validation

All expected values come from the spec or the formula above, never from a decoder. As a syntax check only, an independent decoder (Pillow 12.3) was used three ways:

• It decoded every progressive fixture, including the ZRL refinement case, to exactly the same pixels as its baseline twin (difference 0).
• The quality scaling above matched that library's own tables at qualities 1, 10, 25, 50, 75, 90 and 100.
• At quality 90 with 4:2:0, gradients like these round-trip with a mean error of about 0.8/255, against the 3/255 limit.

scripts/ci: CI: PASS before and after the rebase on origin/dev (1808edc). This PR is 247 lines in 4 files.

Refs #562

Belt-Lead: act-1a1511a1174-tlch
Feature: image-codecs
Slice: IC-a
Fast-Check: pass
Head: a0d47bd

---

[github-actions]

Fast lane: merged into dev (249 lines, 4 files, Fast-Check marker).