Sylpheed/docs/re/ui-title-paint-order-capture.md

The title screen's paint order, measured from the guest's draw submissions

Status: ✅ CONFIRMED — the order in which the running game paints the title screen's sprites, read off the guest's own draw stream. 🟡 the rule that produces that order is still unknown; four candidate rules are refuted below. ❔ two full-screen layers and one rotated effect pair are unidentified.

This closes the measurement gap named in BACKLOG.md — "the next step is no longer static: it is a per-draw capture of the title screen showing the order the game submits" — and it does so without the guest-code read that entry expected.

What it took: a third RE hook in Canary

Neither existing hook can answer this (BACKLOG.md has the detail): log_draws de-dups by vertex declaration, so a screen's sprites collapse into one record, and the F10 ship capture drops any draw without an f32x3 position stream — which is every UI quad. So log_ui_draws was added on the Canary branch auto/re-ui-draw-order:

• F10 arms it (shared with the ship capture; log_ui_draws is what separates them), and it writes xenia_re_ui_draws_NN.log — every draw of the next N frames, in submission order, undeduplicated;
• --ui_draw_capture_frames / --ui_draw_capture_max bound it;
• each record carries the primitive type, index count, VS/PS hashes, every bound texture's base and dimensions, and the quad's vertex positions.

Two traps were paid for on the way, both worth stating because both produced confident nonsense rather than an error:

• counter_ is not a frame number. It looks like one — the VdSwap packet increments it — but GraphicsSystem::MarkVblank() increments it too, from the vblank thread. A capture bounded by it ended after one draw, having "covered 6 frames", because six vblanks landed between two draws of the same frame. The hook now counts VdSwap packets itself.
• screenshot was grabbing a 10-pixel sliver. The app owns two windows of class xenia_canary — the 1280×745 game window and a 10×10 helper — and the wrapper took the first match from xwininfo -root -children, which lists only direct children of the root and so never contained the real, reparented game window at all. Every screen classification for a whole session was noise from a 10×10 image; the movie-skip heuristic fired on it repeatedly and drove the guest into a save-data probe that crashed it. Fixed to take the largest xenia_canary window from the full tree.

The measurement

Boot to the title screen, arm the capture, decode the quads. The title screen's steady state is 11 draws per frame, every frame, identically:

draw	what
0	untextured full-screen triangle (VS 0A6D1DD7, PS 2E372EA2), positions already in pixels
1–8	the UI sprite shader (PS E59B2B3D), quad lists, positions in NDC
9–10	two full-screen triangles sampling a 1×1 texture — a fade/tint pass

Sprite draws sample 1280×768 texture pages (fmt=6, i.e. 720 rounded up to 64), several pages alive at once — so the bound texture does not name the sprite. The quad's size does: the disc's decoded sprites have near-unique dimensions. tools/re-capture/ui_draw_order.py converts NDC → pixels and matches each quad against sylpheed-cli pak textures output.

Frame 2 of captures/title-screen-draw-order.log, in submission order (full table):

draw 1  (   0,   0) 1280x720   full-screen layer
draw 2  (-147,-209)  883x1134 ROT   rotated effect  ┐ two rotated quads,
draw 2  ( 525,-292) 1299x1303 ROT   rotated effect  ┘ off-screen extents
draw 3  (   0,   0) 1280x720   full-screen layer
draw 4  (  64, 115) 1133x281   ptlogo_back2eff / ptlogo_back2eff5   [22 or 24]
draw 4  (  70, 126) 1120x263   ptlogo_back2                         [23]
draw 5  ( 186, 191)  915x115   ptlogo1                              [0]
draw 5  (1126, 378)   38x18    ptlogo_tm                            [11]
draw 6  ( 134, 306)  992x104   ptlogo2                              [1]
draw 7  ( 294, 655)  691x18    ptcopyright                          [28]
draw 7  ( 384, 551)  512x50    ptbtn00        ← build 2, not build 7
draw 8  ( 371, 536)  538x76    ptbtn00f       ← build 2, not build 7

Bracketed numbers are GP_TITLE.pak build 7's declaration indices (sylpheed-cli screen info … --build 7). Sizes match to ≤4 px; the quads run a few pixels under their sprite (919×113 → 915×115), whose cause is unmeasured — a scale just under 1, or a UV inset.

Two things fall out of it immediately

1. Declaration order is refuted per-draw, not just by a framebuffer. The background paints first and ptlogo1/ptlogo2 (elements 0 and 1) paint after elements 22/23 — an order no reading of the table produces. Reverse order dies with it: ptcopyright (28) is second-to-last, not first.

2. The visible title screen is more than one bundle. ptbtn00 (513×50, the PRESS Ⓐ BUTTON plate) and ptbtn00f (537×76) are not among build 7's 30 elements — they are the whole of GP_TITLE.pak build 2 (and its language twin build 3), a one-element bundle: ptbtn00.rat, kind 0x73002, focus-linked to ptbtn00f.rat, resting at (383,550). The capture puts that quad at (384,551), which is a one-pixel agreement between a static placement decode and the running game, on a bundle nothing had checked before. So the screen the player sees is a composite of at least two builds, and no single build's element table can be the paint order, whatever its order. Any rule has to sequence bundles as well as elements inside one.

This also independently corroborates the pivot fix landed earlier the same day: the full-screen layers cover exactly (0,0)–(1280,720), which is where pivot-anchored scaling puts ptbase2 at 200 % and not where corner-anchored scaling puts it.

Refuted, with this capture as the ground truth

• declaration order and reverse declaration order — above.
• keyframe start time — ptlogo1 (first keyframe t=26) paints after ptlogo_back2 (t=112).
• resting-keyframe time — same pair: rest t=42 after rest t=126.
• the placement-region order and the RATC child order were already refuted statically (BACKLOG.md); nothing here rehabilitates them.

Not settled

• The rule. Nothing yet found in the bundle orders these draws. What has not been checked against this new ground truth: a depth/priority field inside the per-element .rat record (as opposed to the 60-byte declaration entry, which was dumped and carries no such field), and whatever sequences the two bundles.
• Three of the twelve quads are unidentified: the two full-screen layers (one is ptbase2 at 200 %, the other is pteff04 or a PRMD; the capture cannot tell them apart because both are full-screen and both sample a page) and the rotated pair, whose 883×1134 and 1299×1303 bounding boxes match no sprite because the quads are rotated.
• One screen only. This is GP_TITLE. A rule needs a second screen with a background and overlapping foreground; the same hook now makes that a ten-minute run rather than a static argument.

---

Correction (same day, next iteration): the running screen is build 4, not build 7

Everything above was matched against GP_TITLE.pak build 7, because that is the largest build and what screen info defaults to. The game does not run it.

screen info --geometry prints each element's decoded sprite size, and the two builds disagree:

sprite	build 7	build 4 (a60fcb85)	captured quad
ptlogo1.t32	822×100	919×113	915×115
ptlogo2.t32	888×93	992×104	992×104
ptlogo_back2.t32	1000×234	1118×262	1120×263

So the quads come from build 4 — a 24-element build — and the earlier index mapping was against the wrong table. The conclusion does not change, only the numbers: the paint order in build 4's declaration indices is

draw 1,2,4   full-screen layers        9 ptbase2 / 10 pteff04 (not separable — see below)
draw 3       two rotated effect quads  ?
draw 5       ptlogo_back2eff(5) 18|20, then ptlogo_back2 19
draw 6       ptlogo1 0, then ptlogo_tm 7
draw 7       ptlogo2 1
draw 8       ptcopyright 21, then ptbtn00        ← ptbtn00 is build 2
draw 9       ptbtn00f                            ← build 2

Declaration order is still refuted (9 first; 0/1 after 18/19), reverse order still refuted (21 second-to-last), and the two-bundle finding is unchanged — build 2 is a one-element bundle holding ptbtn00.rat at rest (383,550), against a captured (384,551).

Reproduced in an independent run (fresh boot, a signed-in profile, different frame numbers): byte-for-byte the same order and the same rects. That is the second observation the CONFIRMED bar asks for.

The batching rule, which is what makes the order readable

Draws carry one or two quads, and a two-quad draw always holds two elements that are adjacent in paint order and share a texture page. Pages alternate (0B2C, 11C3, 0BE3, 11C3, 0B2C), so this is not a sort by page: the engine accumulates quads while the bound page stays the same and flushes when it changes. The paint order is therefore the concatenation of the draws' quads, in vertex order within each draw.

Two more orderings refuted, and one non-place to look

There is no depth in the geometry. Attribute 0 of a UI quad is k_32_32_32_FLOAT — a float3 — so the stream carries a Z. Every Z in the capture is 0.00000, on every vertex of every quad of every frame. The submission order is the whole of the ordering; nothing is left to a depth test.

There is no ordering field in the bundle either. Dumping build 4's whole declaration table word by word (docs/re/captures/title-build4-decl-words.txt): +28 is 0 on all 24 entries, +32 is 0xffffffff (no parent) on all 24, +36 is 0xffffffff except on the three kind = 0x4 repeat instances where it is 0/1, +44 is 0xffffffff on all 24, +56 is 0 on all 24. The placement region's per-group lead word — the one word of that header this project had never read — is 0 for every group. And the region is followed immediately by the RATC child stream (opt + ptlogo1.t32 + T8aD), so there is no table after it to hold a draw order.

So the paint order is not carried by the file. It is produced by the guest's screen code, and settling it means reading that code (the splash path noted in the emulator-era work — sub_821CC7A0, item vtable 0x820b30b4 — draws with exactly the PS hash E59B2B3D this capture sees).

A correction to this entry's own premise

It says the F10 ship capture "explicitly drops UI draws … it requires an f32x3 position attribute, which a 2D quad stream does not have". These quads have one. Armed on the title screen, xenia_ship_capture_01.log holds 1 150 draw records of which 965 are the UI sprite shader (stride=24 prim=13, VS E0BAFB4F).

The new hook is still the right tool, for reasons that are about the record and not the filter: the ship capture writes no frame boundaries, no submission sequence across frames, no bound-texture identity, and de-dups on a key that happens not to collide here only because the game rebuilds its UI vertex buffer every frame. But "it cannot see UI draws" was wrong, and the cost of the new hook was justified with a wrong reason.

---

The refutations, re-checked against the build the game actually runs

The two time-based orderings were first refuted with build 7's numbers, and build 7 is not what runs. Re-checked on build 4, both still fail, and by the same element:

ordering	build 4 values, in the observed paint order	verdict
first keyframe time	bg 0, back2eff5 62, back2 66, logo1 26, tm 98, logo2 26, copyright 138	not monotonic — refuted
resting keyframe time	back2eff5 74 / back2eff 66, back2 80, logo1 42, tm 116, logo2 42, copyright 160	not monotonic — refuted

A new candidate, and its counterexamples. Painter's order by screen depth would be the obvious runtime rule, and sorting the identified sprites by resting Y very nearly reproduces the capture:

sorted by rest Y :  back2eff5 117, back2 126, logo1 186, logo2 315, tm 378, copyright 655
captured order   :  back2eff5,     back2,     logo1,     tm,       logo2,   copyright

— one transposition away. It is still refuted, twice over: ptlogo_tm (y 378) is painted before ptlogo2 (y 315), and the full-screen background is painted first although ptbase2's resting Y (180) would sort it after both back2 elements. Recorded because "nearly sorted by Y" is the kind of result that gets adopted as a rule if only the agreeing five are counted.

A foothold in the guest code, and what it is not

The remaining avenue is the code, and xenia-rs/sylpheed.db is in the container. What is established so far is small, and stated so it is not mistaken for more:

• The item class from the splash-era work is real. Vtable 0x820b30b4 is class ANON_Class_03A0925F with 25 slots, constructed at three sites (sub_823CB2A0, sub_823CB558, sub_823CBB90, each with a single caller).
• RTTI carries no names. All 851 classes in the database are ANON_Class_########; the retail binary has the RTTI structure but not the type names, so nothing can be found by searching for "screen" or "sprite".
• The format tags are immediates, not strings. RATC appears as addis 0x5241 / ori 0x5443 in ~10 functions in the 0x82188–0x821B1 range, always loaded into r4 before a virtual call — e.g. sub_821881D8 does obj->slot1('RATC', &out), then passes the result to sub_82187B78. So the container walk is behind an interface, and a plain string search for .rat / .t32 finds nothing (2 hits disc-wide, both unrelated).
• A dead end worth recording: searching for the declaration table's 60-byte stride (mulli rX, rY, 60) finds 49 sites, and the most promising, sub_821B4498, turns out to be a time-string parser (MM:SS.mmm — it splits on : and . and multiplies by 60, then 100). The guest does not necessarily index that table with a multiply at all.