Sylpheed Disc Atlas

These are not interchangeable, and mixing them up is the single most productive mistake this project has made. They differ in what they hash, and crucially in case sensitivity.

Reading up from the bottom: a domain chain names an asset by string, that string hashes into an archive TOC, the TOC yields a compressed blob, and the blob's first four bytes say which container format it is. Every arrow in this diagram is a hash lookup or a magic-byte test — there are no directory scans.

Two archives hold most of the disc. sound.pak is 1.07 GB across five data segments — more than every other archive combined — and GP_HANGAR_ARSENAL.pak is large for a menu because it is stage-scoped: 168 of its objects are 28 missions × 6 languages, each a complete Hangar configuration.

Each of these starts from something the player can see and ends at bytes on the disc. Where a chain has a hop that looks like it should be direct and is not, that hop is called out — those are where a naive port breaks.

One IDXD object per stage in every language pak. Its StageResource record is the hub: nineteen fields naming the 3D packages, the enemy roster, the routes, the collision set and the localised objective strings.

The Hangar lists 59 weapons; the disc's Weapon table has 131, and the two name sets overlap in zero values. That is not a mismatch to be reconciled — it is the wrong join. An arsenal item names a hardpoint slot, and the slot carries the weapon.

One IDXD object in tables.pak is the whole cue index. Its SOUNDS record is the only record found on the disc that is entirely named fields — 5 798 of them — and the names are the join key. Cue ids are partitioned by family with no overlap.

One manifest record binds four assets per slot. Measured for this page: 104 slots over 101 distinct movies, 99 with a subtitle track, 99 with a voice track, 22 with a telop overlay.

One archive per screen; each RATC bundle is one context × language build. A bundle's declaration table lists every element with its parent and pivot; the placement region right after it gives each element a keyframe group. Both the tutorial pause menu and the title main menu rebuild pixel-accurately from the disc alone.

Two things are not what they look like. The resting position is neither the first nor the last keyframe — it is the plateau. And the declaration table is not the paint order: a per-draw capture of the running title screen paints element 13 first and elements 0 and 1 late. The real order is a second, reordered child array the screen object keeps at runtime — deriving it from the bundle is still open.

545 bytes: a GDHA container, a zlib payload, then a chunk stream. The 54-byte develop blob has the alphabet 0 locked / 2 developable / 4 developed, and is indexed by strings.tbl order — the display order plus the cut items only the localisation file lists. weapon.tbl's id list is not the index space; that it is also 54 long is a coincidence, and the two agree only to index 32.

These are not gaps in the reverse engineering. They are properties of the shipped disc, and each one was found by a check that expected the opposite.

• UI paint order from the bundle. Solved at runtime, not statically — and static is what a port needs.
• The .prt telop format. 22 cutscenes bind one; no parser exists.
• Collision meshes. CollisionSet_S<NN>.bin is documented and unparsed.
• 47 XBG7 resources still miss — mostly pose/proxy composites and damage variants, not a threshold away.
• ~11 GHAD save fields unnamed, and difficulty-versus-stage undecided: three fields hold the value 2.
• The naming of the 2D and Ready Room archives, which is the largest single block of unreachable content on the disc.