Commit both encodings of five renders, for pgb to parse

[turner]

Refs #25.

The two repositories are one contract in two halves — this one produces documents and band payloads, pgb parses them — and nothing checked that they agreed. A wire-format break reached the frontend as an error card at best and a silently wrong picture at worst.

What lands

tests/fixtures/exchange/ — both encodings of all five real subgraphs, out of one render each: the band payload /seqtubemap?format=bands returns, and the document the same request returns without it. 3.35 MB of payload and 3.22 MB of gzipped document, spanning 90 bp to the two regions at pgb's fetch ceiling. The file names are pgb's, so handing them over is a copy rather than a translation.

npm run fixtures:exchange — regenerates them from inputs already committed. No server, no vg, no graph data, no network; about two seconds, and deterministic down to the gzip, so a diff here is always something that moved.

tests/node/exchange-fixtures.test.mjs — the committed payload is the bytes this encoder writes, the committed document is the text this emitter writes, and the committed pair describes one picture, checked through pgb's own parser rather than this repository's idea of one. A sixth test nudges band 0's x1 by a part in 10,000 and asserts the pair check notices, so #25's demonstrated is a standing test rather than a story.

Docs — the fixtures README says what they are, when to re-generate, and that the copy across goes in the same window; docs/band-format.md and the seqtubemap fixtures README point at it.

Verified against pgb's reader

These exact bytes through parseBandPayload.ts and parseBands.ts at pgb fde52e3, over all 101,742 bands: band count, frame, strand names, pclaiScores, ordinates, directions, strand ids and all 1,219 segment boxes equal; abscissae and widths within one Float32 ulp — worst 0.86, 0.86, 0.44, 0.86, 0.86, which reproduces the figures pgb's own fixture notes record. Altering one float fails both sides: this one names the first differing byte, pgb puts the band 16.59 ulp from where the document does.

Against the pairs pgb captured from the live server

Same day, different provenance, and the difference is written down in the README:

• every payload body is byte-identical, as are the header's document dimensions and all 1,219 segment boxes;
• every strand name differs in one way — the live walks append a reference range (…#0[25397889-25398120]) that the 2026-08-27 .gfa subgraphs here predate;
• five strands of 2,880 differ in colour and pclaiScore, having gained a placement on the server since.

What this does not close

#25's remaining criteria are pgb's. Its reader and test exist (pgb#145), but pgb has no CI — no .github/ at all — so "that test runs in pgb's CI" is still open on that side.

Testing

npm test: 132 pass, 0 fail, 2 pre-existing skips. The Python suite is untouched by this change and does not run locally here (fastapi not installed); CI installs it.

🤖 Generated with Claude Code

https://claude.ai/code/session_015wyakF7DEbvddMqaL9kSB6