The subset — measured, guarded, deliberate

Every JPEG 2000 codestream the feeds ship — the golden corpus’ twelve OpenJPEG-encoded fields (GDPS, GEPS, HRDPS, RDPS, REPS, RAQDPS) plus the JasPer-encoded shape a live RDPS CAPE field surfaced the day this decoder went to production — falls inside this shape, and this decoder covers exactly it:

• J2K Part 1 raw codestream (no JP2 container — GRIB embeds bare SOC..EOC)
• one tile covering the whole grid, one tile-part
• one grayscale component, no subsampling — 12/16/20/24-bit unsigned observed; any depth to 28 bits, signed or unsigned, accepted
• reversible 5/3 wavelet, no quantization (QCD style 0; a tile-part QCC restating style-0 quantization for the one component is honoured — the JasPer fields ship one whose exponents really differ from the QCD’s)
• one quality layer, DEFAULT precincts — the implicit 2^15 grid, which is one whole-tile precinct for every gridded field and honestly several for JasPer’s one-row bitmapped fields (813275×1: 25 precincts at r=5) — default codeblock style, 64×64 codeblocks, 5 decomposition levels (any level count in [0, 32] accepted — the loops are generic), no ROI

Loud failure is the design

Everything outside the subset fails with a named UnsupportedJ2kError saying what feature and why: multiple tiles/tile-parts/components/ layers, the 9/7 irrational wavelet, explicit precinct partitions, RGN/COC/POC, main-header QCC, PPM/PPT packed headers, every non-default codeblock style bit by name, SOP/EPH markers, quantization.

Absence of generality is the design: a tripped guard means a feed changed shape, and that deserves a loud failure, not a silent slow path. The guards live in src/codestream.ts; the error vocabulary (UnsupportedJ2kError, J2kFormatError) in src/errors.ts.

One guard is deliberately loose: the three resolution-major progression orders are accepted interchangeably, because with one layer and one component they emit the same resolution-then-precinct packet sequence; the position-major orders (PCRL/CPRL) are accepted only while every resolution has a single precinct, where all five orders coincide.

The JasPer story

The guards earned their keep on adoption day. Most ECCC fields are OpenJPEG-encoded, but the live RDPS smoke surfaced a CAPE field whose COM marker says “Creator: JasPer” — caught by exactly such a loud failure, then extended into the subset deliberately and oracle-verified. It differs in three load-bearing ways from every fixture in the golden corpus:

• Bitmapped one-row geometry: the coded values are flattened to an 813275×1 image (section 6 bitmap; the grid is 1140×1045).
• Multiple precincts: 813275 exceeds the default 2^15 precinct, so the upper resolutions honestly have several precincts (25 at r=5) and the one-packet-per-resolution degeneracy breaks — this fixture is why packets.ts walks the precinct grid.
• Tile-part QCC: quantization is restated in the tile-part header with different subband exponents than the main QCD — mis-honoring it mis-decodes every codeblock — and 24-bit samples, deeper than the 20-bit RAQDPS corpus fixture.

The fixture (rdps-cape-sfc-jasper) is committed beside the golden corpus with its own oracle answers; the full provenance is in grib/test/fixtures/README.md. How it is gated is covered in Two-ring correctness.