[prakashUXtech]

Third part of IC-2 (after #590 and #592). It adds the JPEG fixtures and the baseline JPEGReadTests. Every test is disabled until the task that implements it, following lead.md rule 13. No Sources file changes.

What

Tests/CoreRasterTests/Fixtures/JPEGFixtures.swift builds JPEG streams byte by byte from the specs, and builds their reference pixels from the spec maths:

• Tables. The zigzag order (zigzag) and the K.1/K.2 quantization tables (lumaQ, chromaQ). The K.3 to K.6 Huffman tables (dcLuma, dcChroma, acLuma, acChroma) as DHT carries them, with codes generated by C.1 and C.2 (codes, huffman).
• Entropy coding. Bits writes most-significant bit first, stuffs 0xFF00 and pads with 1-bits. magnitude gives the F.1.2.1 category and extra bits. encodeAC handles run/size, ZRL and EOB.
• Marker segments. segment, jfif (APP0 1.02), exif(orientation, littleEndian:) (APP1 with TIFF IFD0 and tag 0x0112 SHORT, in MM or II form), dqt, dht, sos and header.
• Frame. Size, sampling factors, tables and a quantized block for each (component, block x, block y). mcus, blocks and units give the coding order: MCU-interleaved per A.2.3, or a one-component scan over the component's own block grid per A.2.2.
• Builders.
  • baseline(f, sof:precision:restart:orientation:littleEndian:): a full sequential stream, with DRI and RSTm wrapping 0 to 7 and the DC predictions reset.
  • stub(f, sof:): headers plus a 2-byte scan.
  • split / join: segments before SOS, and the scan, for building damaged cases.
• Expected pixels. seeded(seed) gives deterministic SplitMix64 blocks. idct is A.3.3 in Double, plus 128, clamped. reference(f) applies box chroma upsampling and then JFIF YCbCr to RGB. error(data, f) returns the mean and max error in sRGB levels; alpha must be 1.

JPEGReadTests (all .disabled("waits on IC-6")):

Test	What it checks
infoReadsFrame	JPEGInfo for 4:2:0, a SOF2 frame reports progressive, a side of exactly maxSide is accepted
greyDCOnly	samples are 128 + 2q exactly (within 0.01 level)
zigzag	64 blocks, block k holding only zigzag coefficient k under a flat table; any swapped entry is about 11 levels off. Covers the mutation .zigzag
ycbcrSampling	4:4:4, 4:2:2 and 4:2:0 at 13x11 and 37x21, grey, and SOF1 8-bit decoding the same as SOF0; all within mean 1/255 and max 4/255
restartResetsDC	intervals 1 and 3 on 20 one-block MCUs (RSTm wraps twice), intervals 1, 2 and 4 on 4:2:0, and output equal to the same image with no restarts. Covers the mutation .restartResetsDC
exifOrientation	orientations 6 (MM), 8 (II), 3 and 1 are reported; pixels are not rotated
unsupportedThrows	SOF9, SOF10, SOF3, SOF5 throw unsupported from both info and read; 12-bit and 4-component throw from read
refusesHuge	maxSide+1 on either side and 20,000² throw tooLarge from read before the 2-byte scan matters; info refuses only a side
damagedThrows	no DHT, no DQT, missing DQT 1, no scan, a cut scan, a 16-ones code, sampling 0, an unknown scan component, no SOF, and a segment running past the end throw damaged; no SOI throws notThisFormat

Which task enables what

• IC-6: every test in this PR.
• Next PR: JPEGFuzzTests (IC-6), the progressive tests in JPEGReadTests (IC-7) and JPEGWriteTests (IC-8). It is split out to stay under 400 lines; this PR is 394.

Spec sources and readings

Sources are ITU-T T.81 (A.2, A.3.3, Annex B, C, F.1.2, K.1 to K.6), JFIF 1.02 and EXIF 2.3 (TIFF IFD0, Orientation).

Places where the specs leave a choice, and the reading I picked:

• Upsampling. T.81 does not define chroma upsampling. The references use box replication (sample x·H/Hmax), as the packet's pseudocode says ("box for 4:2:x"). A smoothing upsampler will fail the tolerance on 4:2:x.
• Rounding. The IDCT reference is unrounded Double. The tolerance absorbs a decoder that rounds samples to 8 bits.
• Where 12-bit and CMYK throw. They are asserted on read only. The doc says info throws for frame types it does not read, and these two are not frame types.
• Segment order. EXIF APP1 follows JFIF APP0.

Validation

• I checked the K.5/K.6 value lists for 162 distinct values with counts matching BITS.
• As a syntax check only, an independent JPEG decoder (Pillow 12.3) opened every generated stream, with restarts, EXIF in MM and II, and 4:4:4, 4:2:2 and 4:2:0. All expected values come from the spec maths, not from that decoder. Against the references:
  • Grey and 4:4:4: mean at most 0.46 levels, max at most 1.75.
  • 4:2:x: luma matches within 0.94 levels, so the block order is right. Chroma differs only by the smoothing upsampler that decoder uses.
• scripts/ci: CI: PASS, before and after the rebase on origin/dev.

IC-1 was squash-merged to dev as 13f1006 (#575).

Refs #562

Belt-Lead: act-1a1511a1174-tlch
Feature: image-codecs
Slice: IC-a
Fast-Check: pass
Head: 8723542

---

[github-actions]

Fast lane: merged into dev (394 lines, 2 files, Fast-Check marker).