Round trip for the FS format, the one that goes onto the card

[lgnap]

What

The third of the three formats, and the only one without a round-trip test: FsStoryPackWriter and FsStoryPackReader produce and consume the folder that ends up under .content/ on a device. FsPackRoundTripTest, ten cases, specifications and characterization together, in the mould of BinaryPackRoundTripTest and ArchivePackRoundTripTest.

What survives: version, night mode marker, nodes in order, asset bytes exactly and per node, transitions, the shared action node, control settings; an asset used by several nodes is stored once (header counts and one file each).

What does not, pinned as KNOWN GAP:

• the pack uuid — the writer names the folder by the last eight hex digits of the uuid (the device's convention) and the reader takes the folder name as the uuid, so read back from the folder it wrote, the pack comes out as 56789ABC. LibraryService sidesteps this by renaming to <full uuid>.converted_<millis>, which the second case shows works;
• the node uuids — no room in the 44-byte record; the reader gives node 0 the folder name and invents the rest, differently on every read (the FIXME in FsStoryPackReader);
• a missing audio — the writer substitutes a silent MP3 the device needs, and the reader cannot tell it from a real one.

Two cases pin the bytes on disk against the layout verified on a real card in FORMATS.md §5–6 (#59): the 512-byte ni header field by field, the 44-byte node record and its padding, the 12-byte ri/si entries (000\00000000), li as node indices, .cleartext present, bt absent.

Fixtures

The writer refuses anything but a 320×240 4-bit RLE bitmap and an ID3-less mono MP3, so PackFixtures' pattern bytes cannot be used. The assets go through the project's own ImageConversion.anyToRLECompressedBitmap and AudioConversion.anyToMp3 from a drawn image and a synthesised tone — one MP3 encode per class (~0.5 s), and the two converters exercised on the way. No device data, no third-party content.

TESTING.md coverage map updated with the row. The count row is not touched: it reports the Windows CI figure, which this PR's run will give.

Locally: mvn … test → 340 run, 0 failures (39 FAT32 opt-in skips). git diff --exit-code clean.

Part of lgnap#2 — remaining there: cross-format conversions, PackAssetsCompression, failure paths.

🤖 Generated with Claude Code