OpenMaidEngine/docs/engine-re.md

Native-engine reverse engineering (Ghidra + MCP)

Static RE of the unpacked AGE.EXE engine image, driving Ghidra 12.1.2 via the bethington/ghidra-mcp bridge. This is the home for decompiled native-op findings — the class of logic the scripts call but that lives compiled in the engine (decision→scene, call-script dispatch, op 0x60, the gfx command-buffer). Opcode semantics recovered here also flow into vm-map/opcodes.toml.

Related: docs/scjump-progression.md (the SCJUMP decoder that hit this wall), name-resolution.md §1 (call-script), vm-mapping-plan.md appendix (why the exe is packed + the runtime-dump route).

---

Runbook — the Ghidra + MCP loop

One-time setup (done 2026-07-07):

• MCP server: bethington/ghidra-mcp, cloned to S:\Game Hacking\ghidra-mcp. We used the prebuilt extension GhidraMCP-5.14.2.zip (installed in Ghidra via File > Install Extensions) — this skips the Maven/Java-21 build. The Python bridge runs from a venv (.venv, Python 3.11, pip install .); no uv needed. Registered in Claude Code via .mcp.json at the workspace root: {"mcpServers":{"ghidra":{"command":"S:\\Game Hacking\\ghidra-mcp\\.venv\\Scripts\\bridge-mcp-ghidra.exe","args":["--transport","stdio"]}}}.
• In Ghidra: enable the GhidraMCP plugin (File > Configure) and Tools > GhidraMCP > Start MCP Server (serves http://127.0.0.1:8089/). The bridge talks to that; Claude reaches the bridge over stdio.

Loading the engine image (IMPORTANT — the language gotcha):

• Import age-reimpl/build/engine-dump/range_00400000.bin (the module dump: 2,490,368 bytes, the full 0x400000 module image; VA→file offset = VA − 0x400000).
• Format = Raw Binary, Language = x86:LE:32:default, Image Base = 0x400000. Ghidra's language picker offers x86:LE:32:System Management Mode as the "closest" match — do NOT use it. SMM is a 16-bit segmented (segment:offset) variant for BIOS/SMRAM; it mis-decodes flat 32-bit code (it loaded with addresses like 0000:0000/0025:ffff and produced 0 functions). The plain default variant is correct and yielded 2,721 functions.
• We drove the (re)import over MCP: import_file(language="x86:LE:32:default", compiler_spec="windows", auto_analyze=false) → set_image_base(0x400000) before analysis (so absolute-address refs resolve) → run_analysis.
• Load sanity check (AGF-decoder landmark): at VA 0x474f23, CMP word ptr [ESI + 0x4], 0x4d42 (the BM/BMP-magic check) confirms the image is correctly based + decoded.
• Escalation (unused so far): bin/pe-sieve32.exe /pid <PID> /imp 3 /dmode 3 /dir <out> (run from PowerShell, not Git Bash — it mangles /flags) rebuilds the IAT into a clean PE. Only needed if raw-dump analysis is inadequate; it was fine for reading logic, so we stayed on the raw dump.

---

Master key — the opcode→handler dispatch table (2026-07-07, anchored)

The interpreter dispatches each op via a per-context handler table, fully anchored:

handler(op) = ctx[0x26c93 + op] (word index) = *(ctx + 0x9b24c + op*4) — ctx = the engine context (esi in handlers, thiscall; param_1 in the decompile of the registration routine).

The registration routine FUN_00413860 first fills 0x400 (1024) slots starting at ctx[0x26c93] with a default handler FUN_004162b0 (op 0's slot), then overrides specific opcodes: ctx[0x26c93 + op] = <handler_va>. So opcode = (word_index − 0x26c93). Cross-check: ctx[0x26e3f] = 0x427fb0 (byte offset 0x9b8fc) → op 0x26e3f − 0x26c93 = 0x1ac.

Why this matters: the Kelebek u00XXXXXX opcode names encode handler VAs from Kelebek's build, which drift in ours. This table resolves the real handler for any opcode in our image — the general fix for VA drift project-wide. To find op N's handler: read ctx[0x26c93 + N] from the FUN_00413860 decompile (or *(ctx + 0x9b24c + N*4) at runtime).

Other confirmed engine-context offsets (ctx/esi): +0x53d14 = current gfx-object index; +0x53d88 = per-object cmd-type table (stride 0x78 = 120 bytes); operand-fetch helper = call 0x41b940 (thiscall, ecx=ctx, arg = operand index → returns the operand value); FUN_00415f30(i) = a companion operand accessor.

---

Findings

op 0x1a2 (u00428010) is a GRAPHICS command-buffer op — NOT save, NOT decision→scene (2026-07-07)

The SCJUMP slice assumed u00428010 resolved a decision value to a scene. That premise is wrong, and pinning the real handler via the dispatch table above corrects two layers of confusion:

• VA-drift trap: Kelebek's u00428010 = op 0x1a2. But Kelebek's raw VA 0x428010, in our build, sits inside a different handler 0x427fb0, which is op 0x1ac (per the table: ctx[0x26e3f]=0x427fb0). Op 0x1ac is a save-path op — its handler formats %s\SAVE%2.2d.DAT (format string 0x571e70) and is multi-operand. Reading the raw VA gave the wrong opcode.
• Op 0x1a2's real handler = FUN_0042d360 (= ctx[0x26c93+0x1a2] = ctx[0x26e35]), argc 1. It: sets the current gfx-object cmd-type to 3 (*(ctx+0x53d88 + ctx[0x53d14]*0x78) = 3), fetches operand 1, formats a key with "%c%8.8x" (format string 0x5714e0) of (3, operand), and calls FUN_0042cf70(key, &operand). This is a graphics command-buffer registration op, not save and not scene-load.
• Consequence — the decision→scene premise is discredited. The FIELD snippet lookup(0x5f0ed, 0x62ccf); mov(ptr,1); lookup(0x5f0ed, 0x62ccf); u00428010(ptr) (next op 0x21b, also gfx-family) is a graphics/UI operation, not scene sequencing. So u00428010 does not resolve decision→scene. The real decision→scene mechanism is unidentified — it belongs with the call-script / script-load dispatch (name-resolution.md §1), the next target for this loop (now armed with the dispatch table to resolve the call-script handler directly).

Lesson: never analyze a native op by its Kelebek u00XXXXXX VA directly — always resolve the real handler through the dispatch table (ctx[0x26c93 + op]). The raw VA is off by whole functions.

---

op 0x03 (call-script) is a raw index into the SYS4INI file table — SOLVED (2026-07-07)

The long-deferred call-script <id> registry (name-resolution.md §1) is cracked. Resolved through the dispatch table (op 0x03 → ctx[0x26c93+3] = FUN_0041bc90), then the loader/resolver chain:

• FUN_0041bc90 (handler): fetches operand 1 (the id), bounds-checks call depth (≤ 0x26), pushes a script frame, and calls the loader.
• FUN_0040e980 (loader): opens the resource by id, reads the 0x20-byte SYS4 header, checks magic, allocates per-frame code/local buffers from the header var-counts, reads the bytecode body, and pushes a script frame (stride 0x1e = 30 dwords, indexed by ctx[0x14f45]). Returns to the caller when the callee ends.
• FUN_0044f390 (resolver — the key): record = [ctx+0x414] + id*0x50. The record is exactly the SYS4INI 80-byte layout {name[64], arc_id@0x40, file_number@0x44, offset@0x48, size@0x4c} (count = [ctx+0x40c], archive-name table = [ctx+0x410]). It tries a loose override first (CreateFileA on record.name → the mod/patch hook point), else opens archive [record.arc_id*0x100 + ctx+0x410], SetFilePointer to record.offset, size = record.size. High-byte-tagged ids (id & 0xff000000) select an alternate pack via [ctx+0x3028] — unused by the corpus (0/297 ids carry a high byte).

So call-script <id> = a direct RAW index into the SYS4INI global file table — the same table parse_sys4ini.py reads, but indexed without skipping @ placeholders (13208 records, 2 placeholders). There is no separate on-disk id→code registry; SYS4INI is the registry, and we already had it. Statically confirmed: all 297/297 distinct corpus call-script ids resolve to a .BIN script with a semantically-exact name (0x1ab→ADDITEM, 0x2ae7→MES, 0x143→BUNKI, 0x329d→CALCREVISE, 0x2add→CALCBTPARAM), 0 out-of-range, 0 pack-branch. Tooling: parse_sys4ini.py emits build/callscript-names.json (id→name); sys4load annotates call-script 0x1ab =ADDITEM.BIN; the whole build/disasm/*.asm call graph now reads by name. See name-resolution.md §1.

Companion — op 0x8f (call) is INTRA-script, not cross-script. Its handler FUN_0041fba0 sets [frame PC @+0x53d2c] = [frame codebase @+0x53d28] + operand*4 and pushes a return address on the per-frame return stack ([ctx+0x552e8]/[ctx+0x55248]). The operand is a code offset within the current script (matches header table T3, tag 0x8F = local call targets). So 0x8f is a local JSR; only 0x03 loads another script.

Follow-up (functional): the C# VM still stubs call-script. With the id→resource mapping now known, it can be implemented for real (load the target .BIN from the archive via the SYS4INI record, push a frame, run, return) — the unlock for subroutine-using scripts and, via the same path, decision→scene (scenes are just SCxxxx.BIN records loaded by their SYS4INI index).

---

op 0x215 (query-gfx-object?) is a native command-buffer op — settles the render drift as (b) (2026-07-07)

This is the canonical account of the background/sprite "drift" bug (background pinned off-centre / bottom-right, rest grey — Screenshot 2026-07-06 211353.png). It supersedes the earlier "drift = state-divergence, seed state and it's fixed" conclusion in docs/phase-a-slice-plan.md and the status memory, which are corrected to point here.

Resolved via the dispatch table (ctx[0x26c93 + 0x215]): the registration routine FUN_00413860 stores [ESI + 0x9baa0] = 0x42a0b0, so op 0x215's real handler is FUN_0042a0b0. (Kelebek's 0x421160 is VA-drift — it lands inside the unrelated FUN_00421090. Same lesson as 0x1a2: never trust a Kelebek raw VA.)

FUN_0042a0b0(ctx) does exactly two things:

1. *(ctx + 0x53d88 + ctx[0x53d14]*0x78) = 5 — writes cmd-type 5 into the current gfx-object record. A command-buffer registration side-effect, directly parallel to op 0x1a2 (FUN_0042d360) writing cmd-type 3. So 0x215 is part of the gfx command-buffer subsystem, not a pure query.
2. out = FUN_0047f280(FUN_0041b940(2)) — FUN_0041b940(2) fetches operand 2 (the bytecode handle key); FUN_0047f280 is a std::map::find over an engine-internal associative registry, returning the mapped value or 0xffffffff (not-found); FUN_00425fb0(1, out) writes it to operand 1. That registry is populated by sibling gfx ops — op 0x1a2's handler builds a "%c%8.8x" key and calls FUN_0042cf70, an open-addressing hash insert into the same kind of store.

(a) vs (b) — the verdict is (b). The value 0x215 returns is native command-buffer state: "has a gfx object already been registered under this handle?" (≥0 = existing → use its slot; -1 = new). That state lives in the engine's own registry, maintained by the gfx ops, not in the VM global bank. So seeding story-state globals cannot reproduce it — the drift is not the Phase-B state-divergence problem. Stubbing 0x215 returns a constant → label_12649's slot-select always takes one branch → every draw collapses onto slot 0 → the anchor-preserve math measures foreign-sized textures → cumulative drift.

Why the prior "state-divergence" conclusion was wrong. It was grounded in capture_gfx_objects.py, which polled the object-record array ([esi+0x53d64]) at ~2/s and saw only 3 persistent UI objects, "0 CG objects." But (i) the branch is driven by the map lookup (a different structure the poll never observed), and (ii) command-buffer records are transient — a 2/s poll can't prove CG records weren't used. Absence in that capture ≠ absence of the native path.

The fix is tractable and Frida-free. (b) does not mean an opaque native state machine. The subsystem is a modelable data structure: an object-record array (slot / geometry / cmd-type per object) plus a handle→object registry (a hash map). The gfx ops are inserts/queries/writes against these, and the inserts are bytecode-driven — so a faithful host-side model, with the gfx ops (0x1a2, 0x215, and the 0x212–0x21a family) executed instead of stubbed, rebuilds the state from the same scripts. The opcode- level summary lives in vm-map/opcodes.toml op 0x215.

The query registry is SEPARATE from the geometry object store (2026-07-07) — the retained-mode "2nd CG off-screen" fix

Modelling the gfx ops (above) exposed a subtle but decisive point that the first retained-mode implementation got wrong. There are two distinct native structures, and they must stay distinct:

1. The op-0x215 query registry — a std::map<handle,value> populated ONLY by op 0x1a2 (FUN_0042cf70 hash insert; native stores map[handle] = handle). 0x215 does map.find(handle) → the found value (which equals the handle, and for small system/UI handles doubles as their surface slot) or 0xffffffff = -1.
2. The geometry object store — per-handle V18/V24/V16c/color/draw-bind, touched lazily by the geometry SET ops and draw-texture (gfx_object_get_or_create). This feeds the compositor.

The first GfxState conflated them: GetOrCreate (called by every geometry/draw op) also assigned a fabricated per-object slot via an AcquireSlot() allocator, and QuerySlot (op 0x215) returned it. That is a fiction with no basis in the engine — the native 0x215 never allocates a slot.

Consequence, traced end-to-end in SC0000 label_12649 (the CG-load subroutine): a CG handle (0xcb2a = INIT2's G[0x62456], idx 1) is never 0x1a2-registered. Real engine → 0x215 returns -1 → the fresh branch runs → anchor comes from the INIT2 arrays (G[0x62469+idx]=400, G[0x6247d+idx]=600) → dst = anchor − (w/2, h) = (0,0). Correct. But with the fabricated allocator the second pass over the same handle found it "existing" (slot 4) → the existing branch ran get-texture-size(4) on a slot whose surface was never loaded (the bytecode's own slot table rec[s3]/G[0x3239] gave slot 0) → size 0 → anchor = pos(0,0) + 0 → dst = (0−400, 0−600) = (−400,−600) — the CG rendered off-screen. This is the bug that had been mis-attributed to "geometry accumulation / drift" several times.

Fix (branch feat/gfx-command-buffer): GfxState keeps a separate _registry (a HashSet<long>) populated only by Register(handle) (op 0x1a2); QuerySlot returns handle if registered else -1, and no longer consults the geometry store or invents slots. Verified: Age.Cli gfx --boot SC0000.BIN → all event CGs dst=(0,0), zero (−400,−600) draws; Godot --boot --shot pages 1/2/4 render the opening event CGs full-screen; engine 44/44; sweep parity 284 exit / 13 STEP-LIMIT unchanged.

Note a second, still-latent gap this uncovered: label_125bd (which fills rec[s3]/G[0x3239] with the per-object slots 4..13, called at SC0000 0x50f) does not execute in a cold single-scene run — the scene coroutine framework (ops 0x7b/0x140 + the G[0xaba5c]==1 re-entry gate) routes cold flow past it, so every fresh CG is assigned slot 0. It doesn't break the opening (one full-screen CG shown at a time, so sharing slot 0 is harmless and the fresh-branch geometry is correct regardless), but a scene with several simultaneous distinct-slot objects would need the setup to run. Tracked as the scene-coroutine work, separate from this fix.

⇒ Scene-coroutine framework — the fix for the slot-0 collapse (2026-07-09, RE head-start). To run label_125bd (and every scene's slot-table setup) generally, the scene-coroutine framework must work. Ops (handlers via dispatch table ctx[0x26c93+op]):

• 0x7b (FUN_0041ebf0) — yield-state save: writes op1→ctx[0x6da88 + ctxidx*4], op2→ ctx[0x6db28 + ctxidx*4] (ctxidx = ctx[0x53d14], the coroutine/script-context index).
• 0x7c (FUN_00417cb0) — yield/resume: requires run-state bit 0x2000000 set (ctx[0x6dbc8]); restores PC = ctx[0x53d28] + ctx[0x6dbcc]*4, clears the run-state (ctx+0xa0ce4 &= ~0x2000000), resets input/line state. This is the coroutine RESUME.
• 0x140 (LAB_004299c0) — the LABEL/yield op (u0041F9C0, "LABEL" "J" in SC0000 0x46d); a LAB not a FUN, so decompile the actual target next. Central to the re-entry.
• Gate: G[0xaba5c] (==1 re-entry gate at SC0000 0x450); the scene runs as a coroutine that yields and re-enters, and label_125bd (slot setup) is reached only on the correct pass. G[0xaba5c] writers = DEBUGADV/SC0000/SC0010; compared against [0,1]. Context records are the 0x78-byte coroutine records at ctx+0x53d14/0x53d88 (see §"Frame cadence"). NEXT: finish RE (0x140 target, the gate branch semantics), then implement so cold single-scene runs execute the scene's setup — removing the need for the manual --seed slot-table unblock.

Revealed issue (2026-07-08, post slot-fix): a magic-circle effect persists across the scene transition (screenshot: opening ritual circle still overlaid on the arena BG). A retained object not released at the phase change — plausibly the same coroutine/lifecycle gap (scene-phase cleanup), or a separate release/clear op. Verify once the coroutine framework runs.

gfx command-buffer — op contract table (2026-07-07, full family reversed)

Every gfx op shares one shape: write a cmd-type into the current object record (*(ctx + 0x53d88 + ctx[0x53d14]*0x78) = <cmd>), fetch operands via FUN_0041b940(i) (1-based; docs = the 0x1a2 variant uses FUN_00415f30), then either SET object fields (call a native worker FUN_0047xxxx) or QUERY object fields (write results back to output operands via FUN_00425fb0(i, val)). Handlers resolved through the dispatch table (ctx[0x26c93+op]); all renamed in the Ghidra project gfx_op_0x<op>_<role>.

op	handler	cmd	dir	argc	contract
0x1a2	0x42d360	3	set	1	registry insert: key "%c%8.8x"(3, operand-desc) → FUN_0042cf70
0x1f7	0x422270	5	erase	2	registry erase (teardown, NOT create): op2>1 → gfx_registry_erase_range(op1,op2) erases [op1,op1+op2), else gfx_registry_erase(op1). Objects are created lazily by the geometry SET ops.
0x1fa	0x4224a0	3	set	1	release element [ctx+0x52bd4 + op1*4] (vtbl free) + FUN_00474e40(op1)
0x1ff	0x4227b0	9	set	4	3 int→float params on obj op1 → FUN_0047e800(op1,f2,f3,f4)
0x202	0x4228d0	0xb	set	5	blit obj op1 with (op2,op3) + packed ARGB from op4(alpha)/op5(color) → FUN_0047ea00
0x203	0x4229a0	9	set	4	draw obj op1 with op2 + packed color(op3/op4) → FUN_0047e9b0
0x212	0x4230c0	5	set	2	obj[ctx+0x14d54 + op1*4] -> +0x64 = op2
0x213	0x423110	7	set	3	obj[0x14d54+op1*4] -> +0x68 = op2 ; +0x6c = op3 (an (x,y) pair)
0x215	0x42a0b0	5	query	2	registry find(op2 handle) → op1 (value / 0xffffffff). Drives slot-select.
0x216	0x42a0f0	5	query	2	read [ctx+0x46d14 + op2*0x14] → op1
0x217	0x4231b0	9	set	4	3 int→float on obj op1 → FUN_0047e960 (SETS a geom 3-vector)
0x218	0x42a130	9	query	4	FUN_0047f360(obj op1) → op2,op3,op4 (GETS a geom 3-vector)
0x219	0x423240	9	set	4	3 int→float on obj op1 → FUN_0047e910 (SETS a geom 3-vector)
0x21a	0x42a1b0	9	query	4	FUN_0047f2e0(obj op1) → op2,op3,op4 (GETS a geom 3-vector)

label_12649 correlation (the drift chain, confirmed). The recurring idiom is:

query-gfx-object? (G 0x62452) (G 0x6245X)   ; 0x215: handle G[0x6245X] -> working slot G[0x62452]
ui-elem?          (G 0x6245X) 0xa           ; 0x1f7: select that element
ui-clear?         (G 0x62452)               ; 0x1fa: clear the slot

G[0x62452] is the working slot; G[0x6245X] are per-object handles (the 0x62455[idx] family: 0x62456/7/8/a/b/c). The geometry ops move two per-object 3-vectors between object records and globals:

• 0x217 SET anchor-vector G[0x6249b/c/d] into the object; 0x218 GET it back out.
• 0x21a GET position-vector into G[0x62498/9/a]. These get-vectors are exactly the inputs to the anchor-preserve math (docs/superpowers/specs/2026-07-06-a2b-graphics-geometry-design.md: G[0x62498] = G[0x6249b] − w/2, foot-anchor at G[0x6249c]). So the drift has two stubbed drivers, not one: 0x215 (wrong slot → collapse to slot 0) and 0x218/0x21a (stale geometry vectors → the anchor math reads garbage). Both read object state the SET ops (0x217/0x219/0x212/0x213) wrote — all bytecode-driven, all host-modelable.

Model implication for the host-side reimplementation (Phase 2 input). The subsystem is a set of per-object records keyed by handle, carrying: a slot (from the 0x215 registry), a position 3-vector (0x21a get / a matching set), and an anchor 3-vector (0x218 get / 0x217 set), plus color/blit params (0x202/0x203). The native workers (FUN_0047xxxx = the DirectDraw/surface layer) need not be modelled — only the object-record data model, so the QUERY ops return what the SET ops stored. That makes 0x215/0x216/0x218/0x21a return correct values and the existing bytecode geometry math produces correct dst/w/h. Ancillary per-object tables observed: ctx+0x14d54 (obj pointers, fields +0x64/ +0x68/+0x6c), ctx+0x46d14 (stride 0x14), ctx+0x52bd4 (element pointers), plus the 0x408 registry.

Worker functions decoded + annotated in the Ghidra project (2026-07-07): gfx_registry_erase(0x47d850), gfx_registry_erase_range(0x47d8b0), gfx_object_get_or_create(0x47ddb0, inserts a zeroed default via gfx_object_init_default@0x472810), the setters gfx_set_vec18/24/16c(0x47e960/e910/e800), the getters gfx_get_vec18/24(0x47f360/f2e0).

The 0x21c–0x243 sprite transform / ANIMATION cluster (2026-07-07, recon — implementation pending)

The scene-completeness tracker (tools/scene_opcode_coverage.py) flagged a dense band of GAP ops in 0x21c–0x243 (+ 0x2bd/0x2bf) — the largest remaining rendering unknown in SC0000 (e.g. 0x220×66, 0x22f×34, 0x228×33, 0x21e×25 static sites). Resolving every one through the dispatch table (ctx[0x26c93+op], read from FUN_00413860) shows it is one coherent subsystem: sprite transform + animation/tween — and two members were already named in prior RE (0x234 gfx_op_0x234_anim_start, 0x238 gfx_op_0x238_set_anim_clock). Kelebek VAs drift here as everywhere (op 0x220 real handler is 0x4234e0, not Kelebek's 0x4215D0). Op → real handler map:

op	handler	op	handler	op	handler
0x21c	0x417520 (417xxx trivial)	0x229	0x423700	0x236	0x423ee0
0x21d	0x423310	0x22a	0x4237b0	0x237	0x4240a0
0x21e	0x423350 ✎ set_transform3_norm	0x22b	0x423850	0x238	anim_start's clock ✎
0x21f	0x423410	0x22c	0x423900	0x239	0x424120
0x220	0x4234e0 ✎ set_transform3_abs	0x22d	0x423990	0x23a	0x42a440
0x221	0x423590	0x22e	0x423a40	0x23b	0x424190
0x222	0x4235e0	0x22f	0x423b00	0x23c	0x417580 (417xxx)
0x223	0x423620	0x230	0x423ba0	0x23d	0x4175c0 (417xxx)
0x224	0x417550 (417xxx)	0x231	0x423be0	0x23e	0x42a4a0
0x225	0x4236a0	0x232	0x423c30	0x23f	0x42a520
0x226	0x42a230	0x233	0x423cf0	0x240	0x4245f0
0x227	0x42a2e0	0x234	anim_start ✎	0x241	0x4247e0
0x228	0x42a3a0	0x235	0x423e40	0x242	0x4249d0
				0x243	0x4182d0 (417xxx)

(0x2bd→0x4251c0, 0x2bf→0x425240. The handful of 0x417xxx handlers are trivial/marker-shaped — the default-handler neighbourhood — and are almost certainly no-ops or arg-poppers; triage before modelling.)

Contract (decoded, representative ops 0x220/0x21e, both argc 6, annotated in Ghidra): same shape as the geometry family — write cmd-type 0xd into the current object record, fetch operands 1..6, call a transform worker with (int op1=handle, int op2, int op3, float op4, float op5, float op6). 0x220 uses raw floats (worker 0x47ecc0); 0x21e normalizes the 3 floats by /_DAT_00571c28 (runtime-init divisor, static 0) so operand 0x64=100 → a fraction → scale/percentage (worker gfx_anim_set_channel@0x47eaa0). The worker calls the SAME gfx_object_get_or_create our GfxState already models, then arms an animation channel on the object record: obj+0x3c = op2, obj+0x50 = op3, obj+0x68 = 1 (enable), obj+0xac = vec3(op4,op5,op6) (the transform target), and raises global dirty flags ctx+0xb558/+0xb560. Corpus idiom: 0x220 (handle=0xcb20+k) 800 500 0 0 0 (size a CG object), 0x21e (handle) (val) 100 100 100 100 (scale/color channels). 0x234 anim_start + 0x238 set_anim_clock imply a per-frame clock that interpolates these targets over time — i.e. this is what makes AE* fades/effects animate rather than snap.

Model implication (Phase-2 input, mirrors the geometry family): the DirectDraw workers need NOT be modelled — extend the host GfxState object with the transform/anim fields (a transform vec3 target + the two scalar params + enable + an animation clock), have the SET ops (0x21e/0x220/0x234/0x238/…) write them and the compositor apply the transform per-frame, stepping the clock on anim_start/set_anim_clock. This is a spec/plan-worthy chunk (~18 effectful handlers + workers 0x47eaa0/0x47ecc0 + the per-frame stepping); the op map above is the de-risked starting point. tools/scene_opcode_coverage.py SC0000 measures the GAP shrink as each lands.

anim_start/set_anim_clock decoded + opening confirmed (2026-07-07, animation-slice Task 1)

Decoding the two already-named clock/start ops (dispatch table → 0x234@0x00423da0, 0x238@0x004240e0; both annotated) and grepping the SC0000 opening settles the animation model and confirms the opening exercises it:

• 0x238 set_anim_clock (argc 1, cmd-type 3): ctx+0x51b78 = 0 (elapsed), ctx+0x51b7c = operand1 (total duration). A GLOBAL, NON-BLOCKING clock — not per-object. The op only configures the clock; it does not loop/wait. The native render loop advances this clock each frame and interpolates all animating objects. Its own plate comment states the payoff: "our port can drive animation in the host's per-frame loop while the VM is parked at wait-for-input; no VM/host frame-lockstep." → validates the wall-clock-tween architecture directly. SC0000: set-anim-clock(G[0x624bb]) @0x123bd, set-anim-clock(0x190=400) @0x13858.
• 0x234 anim_start (argc 5, cmd-type 0xb): gfx_anim_start(op1=handle, op2, (float)op3, (float)op4, (float)op5). op2 = this object's animation duration (label_1235a maxes the per-object durations into the global clock); (op3,op4,op5) = the target transform vec3 the object animates toward. SC0000 opening @0x00daf: anim_start(G[0x62457], 0x2328, 0, 0, local2) on the INIT2 CG handles, then call label_1235a.
• The opening path uses the whole subsystem, early. 0x21e/0x220 transform-sets fire from 0x00f73 onward (0x21e (G[0x6245b]) 0 0x12c l0 l1 0x64, 0x220 (G[0x62457]) 0x96 0x3e8 l1 l3 0), on the same INIT2 CG handles (G[0x62457],0x6245b,0x6245c) — this is the opening, not battle/debug. So the slice's ops are real and verifiable on screen.

Corrected host model (supersedes the "per-object clock" wording above):

• Global clock (from 0x238): one AnimClockDurationTicks + a generation/reset marker the host watches to reset its wall-clock elapsed to 0. The host tweens all armed objects over this duration.
• Per-object (from 0x21e/0x220 = set transform directly; 0x234 = animate toward a target): the object's transform target vec3 + the two scalar params + enable + a per-object generation (bumped by anim_start).
• Residual (empirical, Task 6): which vec3 component is opacity vs scale vs position lives in the DirectDraw draw-worker we deliberately don't model. Determine it empirically from the animating channel + screenshot, not by RE'ing the surface layer.

The opening render path is RETAINED, not immediate-mode (2026-07-08, ground-truth correction)

A working note in the animation slice mis-called the SC0000 opening a set of "immediate-mode slot-0 blits." That was wrong, and it came from trusting our own Age.Cli gfx oracle (which executes our VM and mis-labeled the CG draws as "slot 0"). Verified against native code + the raw bytecode:

• draw-texture (op 0x1fb, handler gfx_op_0x1fb_draw_bind@0x422510) is a RETAINED bind, not a blit. It writes cmd-type 0x11 and calls gfx_object_bind_draw@0x47e870, which on the object keyed by handle (operand 1) sets: flag|=1 (visible), obj+4 = source SLOT index, obj+8..0x14 = source rect, obj+0x24/28/2c = position. Its plate comment (prior RE) already states the key fact: the object stores the slot INDEX — a live ref to surface[slot], resolved each frame at render — NOT a texture snapshot. Objects persist and are composited each frame; this is exactly the surfaces+objects model in "The full gfx render model" above.
• The SC0000 opening is a retained scene of distinct objects, sleep-paced. Raw bytecode: fixed-handle UI objects (0xcf08 slot 3 full-screen, 0xc350 slot 0xe, 0xe678 slot 0xd — a 400×30 element re-bound 20+ times), an animated sprite (draw-texture (G[0x62457]) (G[0x62452]) … (G[0x62498]) (G[0x62499]), computed position), and the CG loader (SC0000 @0x126e1/0x12970): set-texture G[0x62424] → slot G[0x62452], get-texture-size, centre it, then draw-texture (handle = CG_array[G[0x62450]] = INIT2 array G[0x62455..]) slot G[0x62452] …. sleep 0x64/0x3e8/0x2ee sits between steps. So different draws use different handles and per-object working slots — not one slot-0 canvas.
• Why our port still doesn't animate the opening (conclusion unchanged, mechanism corrected): we execute the whole load/draw/sleep sequence instantly — no sleep timing, no per-frame present — so we only ever see the final retained state; the intermediate AE* frames (AE001D→AE002B→AE003B, surface swaps on the working slot between paced frames) never get a frame to display. The fix is frame-pacing (scene-coroutine / sleep 0xc8), a separate subsystem from the transform/alpha channel. Lesson: never characterise the engine's render mechanism from our own VM's oracle output — use native code + raw bytecode.

sleep (op 0xc8) — the frame-pacing primitive (2026-07-08, decoded)

Handler resolved via the dispatch table (ctx[0x26c93+0xc8] = param_1[0x26d5b] in FUN_00413860) → sleep_op_0xc8@0x420ec0 (was LAB_00420ec0; created + annotated). It is NON-BLOCKING:

• It arms a timer — sleep_timer_arm@0x44cff0 on the object at ctx+0x5f304: +8 = 1 (active), +0x14 = (*DAT_0056f3d4)() (start tick — an ms source, timeGetTime/GetTickCount class, same DAT_0056f3d4 the boot uses to seed srand via time/100), +0x18 = duration (operand, min 1). The engine's main loop polls elapsed ≥ duration and resumes the script — rendering continues in the meantime. This is the native confirmation that the engine paces animation in its per-frame loop, not by blocking.
• Operand unit = MILLISECONDS. duration < 10 fast-paths through import [0x56f0b8]; every real scene sleep (100/750/1000 in SC0000) is ≥ 10 → the timer-arm path.
• The handler also writes gfx cmd-type 3 into the current object record (ctx+0x53d88+curidx*0x78) and runs two anti-tamper checks (call [ctx+0x5512c]; a rotate-checksum compare of ctx+0x55120/0x55124; __CxxThrowException on mismatch — integrity work piggybacked on a hot op). Neither is needed by our model.

Port equivalent (implemented): our VM runs on a background thread (like wait-for-input), so blocking that thread for duration ms while the main-thread compositor (Main.Recomposite in _Process) keeps presenting is behaviorally equivalent to the native non-blocking timer. This correctly reproduces the explicit one-shot sleeps (the dramatic 1000/750/200 ms holds). Headless/CLI hosts no-op Sleep (parity). IHost.Sleep(long) + VM case "sleep"; see vm-map/opcodes.toml 0xc8. ⚠ It does NOT make the rapid opening AE* burst animate — execution trace shows the back-to-back set-texture→draw-texture swaps have no sleep/wait/present/ coroutine between them, so bare sleep was never their pacer; what advances that burst is still unknown (an earlier claim that "the opening is sleep-paced" was inherited from this doc and never execution-verified — corrected). Profile the real Godot run (--trace-histogram) to find it.

Related — present-frame (op 0x20c): dispatch param_1[0x26e9f] = gfx_op_0x20c_present_frame → gfx_render_frame@0x4820b0 (buffer flip). Our compositor presents every frame regardless, so 0x20c is a VM no-op (noop_headless=true); the Kelebek label u00416200 was VA-drift. This corrects the earlier open item ("no per-frame present") above — present is host-implicit; only sleep timing was missing.

Frame cadence — the interpreter tick, and why our port "speeds through" (2026-07-08)

Answers the open question the sleep section above left ("what advances the rapid opening burst is still unknown"). The pace is an engine-level execution cadence, not any bytecode primitive. Corroborated in-game by Ctrl fast-forwarding ADV (a speed governor). Motivated by the user's observation that our port visibly speeds through the opening — which contradicted, and correctly overturned, an earlier same-day overclaim that "there is no missing pacer" (that was inferred from headless op-counts, which cannot render).

Confirmed from the engine image (annotated in Ghidra):

• The interpreter is a cooperative one-op-per-tick step, not a run-to-completion loop. adv_interpreter_tick@0x410fb0 (renamed from FUN_00410fb0) executes exactly one opcode per call: op = **(ctx+0x53d2c + curCtx*0x78); if 0 ≤ op ≤ 0x3ff it dispatches (*(ctx+0x9b24c+op*4))() (the handler table = ctx[0x26c93+op]) then advances PC += *(ctx+0x53d88+curCtx*0x78) * 4 (decoded cmd size), else the default handler FUN_004162b0. It also runs the message-skip / click / auto-advance logic each tick (s_set_CancelMesSkipOnClick, s_message_ReadTextSkip, skip bit ctx+0xa0ce4 & 0x8000000) — i.e. the Ctrl fast-forward governor lives at the per-op level, and a click can reposition the PC (skip-to-next).
• Script contexts are coroutine records. curCtx = *(ctx+0x53d14) indexes 0x78-byte records at ctx+0x53d60/ctx+0x53d2c (PC, codebase, cmd-size). The engine multiplexes script "threads." Init/reset = scene_context_init_reset@0x40b3b0 (zeroes 0x53d14 + 0xa0ce4, allocs surfaces ctx+0x52bd4[1000]).
• Advancement is gated by an interpreter run-state flags word ctx+0xa0ce4 (bit1 = sleeping, plus wait/ skip/etc.), read+written by ~40 state functions. sleep_op_0xc8 sets bit1 + arms the ms timer and returns — it does not block. So the outer loop consults 0xa0ce4 to decide whether to step the script this frame.
• Effects are frame-stepped. Screen transitions FUN_0043cdb0 (12 wipe/slide modes) render one frame per step and take a step-count parameter (the natural place a speed multiplier applies); present-frame (0x20c) and the anim clock (0x238) advance per frame. A CG transition therefore spreads over many real frames.
• Timing source = the ms-clock function pointer *DAT_0056f3d4 (timeGetTime-class), used throughout.

Model: single-threaded, vsync-timed frame loop; each frame it steps opcodes until the context yields (sleep armed / wait-for-input 0x72 / active frame-stepped transition/anim / present), renders (gfx_render_frame), waits on the clock, continues. Back-to-back draws inside one page compose into a single frame (fine); the opening's CG-to-CG advances are gated by frame-stepped transitions + sleeps, which spread them over real time.

⚠ Not statically resolvable (honest boundary): the outer frame loop itself is not readable from this dump. adv_interpreter_tick is invoked through a runtime-set mode function pointer (heap/vtable slot) — it has zero static xrefs, and its address bytes (b0 10 41 00) appear nowhere in range_00400000 (0x400000– 0x65ffff). The functions touching the scheduler state (0x53d14, 0xa0ce4) are init/reset, save (context_state_serialize@0x40d320), and op-handlers — never the loop. The "run-until-yield then render" statement above is a reconstruction from those pieces, not a line read from the loop; pinning the actual loop + its exact per-frame step budget / vsync wait needs a live-debugger break (attach + break in the frame loop), or a wider memory dump that includes the mode object.

Port relevance (the speed-through root cause). Our Godot VM runs on a free-running background thread (Task.Run(() => vm.Run()) in Main.cs) with no frame binding — it executes an entire page's ops in microseconds; only WaitForInput and Sleep pause it, and the compositor merely samples GfxState at 60fps. So every no-sleep CG/state advance collapses to its end state → the speed-through. Fix shape: throttle the VM to a bounded wall-clock op rate (see the measured numbers below) via a per-opcode host yield; retire the free-running thread. Spec: docs/superpowers/specs/2026-07-08-frame-stepped-vm-design.md.

Frame cadence — live measurement (2026-07-08, Frida read-only)

The static pass couldn't reach the outer loop, so we measured the running game. Read-only / import-only only (tools/frida/probe_frame_cadence.py, probe_present.py): a plain-JS hook on the proven operand-fetch 0x41b940 (grab ctx + count exec rate) + system-DLL hooks; no engine-code patching. Lesson learned the hard way: a first attempt with a CModule hook on the hyper-hot adv_interpreter_tick crashed the game instantly (bad native callback into the hottest path — not anti-tamper; our other scripts hook engine code via plain JS and survive). Use plain-JS hooks on proven addresses + memory polling.

Findings:

• Execution is rate-limited, not free-running. Normal active rate ≈ 1,788 operand-fetches/sec (peak ~5,796) — far below an unthrottled interpreter (millions/sec), so the engine paces itself. Execution is bursty (parked at wait-for-input prompts, then a bounded burst), confirming per-iteration op-budgeting.
• Fast-forward (Ctrl) scales the rate ~4× (≈7,738/sec avg, peak ~15,572), gated by the engine skip bit ctx+0xa0ce4 & 0x8000000 (set only while fast-forwarding). It runs more ops per unit time — it does not skip content. (Ctrl is ADV-scoped; it does not speed up gameplay/menus.)
• Rendering = Direct3D 9, UNCAPPED. ddraw.dll is not loaded; the game uses d3d9.dll (+ nvd3dum.dll). IDirect3DDevice9::Present (device vtable slot 17, found by scanning ctx for a d3d9-vtable object with a full ~119-method table) fires ~1,908/sec with no vsync; BeginScene/EndScene never fire → a 2D StretchRect-style compositor, not a 3D scene. So there is no fixed display-frame rate; Present rate ≈ op rate (~1 op per present). ⇒ the pacing quantity is the wall-clock op rate, not a per-frame budget.
• Implication for the port: throttle our VM to ~1,800 ops/sec wall-clock (≈30 ops per 60 fps Godot _Process, tunable), ~4× under a future Ctrl multiplier; Godot's 60 fps compositor + wall-clock tweens then show the smoothly-advancing state. This measured mechanism replaces the earlier present-driven guess (present is rare in our path and uncapped natively).

The render drift's SECOND half: missing system-boot state (2026-07-07, resolved)

Implementing the gfx ops (above) was necessary but not sufficient — a cold single-scene run of SC0000 still drifted. Runtime tracing found the CG handle array G[0x62455..0x6245c] was all zeros, so every CG collapsed onto object 0 and its geometry accumulated. Those handles are set by the boot script INIT2 (mov 0x62455=0xcb20 … 0x6245c=0xcbc0), which is call-scripted by the real entrypoint SYSTEM4.BIN (LOADCONFIG → INITCONFIG → INIT2 → LOGO → OP → INIT → TITLE → …). Our harness teleports straight into SC0000, skipping that boot. Fix: run the system-boot state prefix (INITCONFIG/INIT2/INIT, skipping the UI scripts) before the scene — Age.Cli gfx --boot and Godot --boot (both via GameSession). With boot, the CGs de-collapse and render correctly (screenshot-confirmed). This is the synthesis of the old (a)-vs-(b) debate: the drift needed BOTH the native gfx ops (b) AND boot state (a) — specifically INIT2's handle array, never before identified (it is not a story flag). Note two distinct boots: our Phase-B --boot runs the data *INIT scripts (skills/items/…); this is the system boot (SYSTEM4 prefix) — a "full boot" should run both. Residual: the AE* fade/flash effects still draw opaque (alpha/blend deferred — Phase 2 scope), and some object-slot CGs start with a zero anchor (cold gfx objects vs the real game's warm ones; default object geometry is confirmed (0,0) in gfx_object_init_default, so it is not a missing-default bug). See docs/phase-a-slice-plan.md A2b-Geometry.

The gfx animation/effects subsystem — the AE* fades (2026-07-07)

The AE* flash/glow effects (and sprite motion) are a native time-animated retained render loop, not per-frame bytecode. Reversed + annotated in Ghidra:

• Retained objects carry animation state: an active flag (obj+0 bit 4), a progress counter (obj+0x214, starts 0), a duration (obj+0x228), and a target vector (obj+0x244/248/24c).
• gfx_anim_start (0x47f060, worker for op 0x234) configures a per-object animation: sets the flag, resets progress, stores duration + target. Op 0x1fd (gfx_op_0x1fd_set_vec_scaled) sets a scaled 3-vector.
• Op 0x238 (gfx_op_0x238_set_anim_clock) sets a global animation clock, non-blocking: ctx+0x51b78 = 0 (elapsed), ctx+0x51b7c = duration (the max per-object duration; SC0000 label_1235a maxes a table to compute it). It does not loop/wait.
• Frame model: the bytecode does configure anims (0x234/0x1fd) → set clock (0x238) → show-text → wait-for-input and continues; the native render loop advances the clock + per-object progress each frame, interpolates, composites, presents. Render/present family nearby: 0x243/0x20c/0x21c/0x224 (u004162xx, not yet fully RE'd). ⇒ the port can drive animation in the HOST per-frame loop while the VM is parked at wait-for-input — no blocking present op, no VM/host frame-lockstep (the answer to the "frame loop" question).

Consequence: reproducing the fades needs a retained per-frame animated compositor — the current immediate-mode permanent canvas can neither fade nor clear. Design: docs/superpowers/specs/2026-07-07- animated-compositor-design.md.

The full gfx render model — surfaces + objects + composite (2026-07-07)

Reversed the create/set/draw-texture handlers + the render loop (all annotated in Ghidra). This is the canonical model (an earlier flat "draw layers to one screen" attempt was WRONG — it had no surface concept and snapshotted textures at draw time; symptoms: alternating grey, glow over backgrounds, vanishing sprites).

Two distinct stores:

• Surfaces — image buffers at ctx+0x52bd4[slot], indexed by slot. gfx_op_0x1f8_create_surface (0x4222d0) allocates a blank one (releasing any old); gfx_op_0x1f9_load_surface (0x422360, op 0x1f9 set-texture) resolves resId via the SYS4INI resolver (FUN_0044f390) and loads the file into the slot's surface with a colorkey/chromakey (op arg 3 — never modelled before), also releasing the old surface. A surface persists at its slot until the next set-texture overwrites it.
• Objects — the ctx+0x408 registry, keyed by handle (a std::map). gfx_op_0x1fb_draw_bind (0x422510, op 0x1fb draw-texture) → gfx_object_bind_draw (0x47e870): sets the object's source slot (obj+4), source rect (obj+8..0x14 = left,top,right,bottom), position (obj+0x24/28/2c = V24), and the visible flag (bit 0). The object references its surface by slot index, live (re-resolved each frame), NOT a snapshot. Objects also carry anchor V18 (obj+0x18), animation (flag bit 2 + progress obj+0x214 / duration obj+0x228 / target obj+0x244..), and color/alpha (0x202/0x203).

Render frame — gfx_render_frame (0x4820b0), driven by op 0x20c present (gfx_op_0x20c_present_frame 0x4174a0, which also updates the frame timer ctx+0x51b64/68): iterate the object registry in ascending handle order — that IS the z-order (lower handle behind, higher on top; std::map key order). For each object with visible bit 0, gfx_object_composite (0x47f650) computes its transform from geometry, applies the animation interpolation if bit 2 is set, and blits surface[obj.slot] with alpha/colorkey. Then swap buffers (present). Slot 0 is NOT special — a normal slot; several objects may share one surface.

⇒ Faithful port: a SurfaceStore (slot → {image, colorkey}, from create/set-texture) + an ObjectStore (handle → {slot, srcRect, position, anchor, scale, anim, alpha, visible}, from draw-texture + the gfx ops) + a host per-frame compositor that draws visible objects in ascending-handle order from their live surface, interpolating animations by elapsed time. No VM/host lockstep (op 0x238 clock is non-blocking; animations play during the wait-for-input park). Open detail for implementation: the exact scale/transform math in gfx_object_composite (FUN_00472f00/FUN_00473ed0).

Blend & transparency — colorkey + 0x202/0x203 color/alpha (2026-07-08)

Reversed for graphics slice A (spec docs/superpowers/specs/2026-07-08-blend-transparency-design.md; Ghidra functions renamed + plate-commented, saved).

• Colorkey format (gfx_op_0x1f9_load_surface 0x422360): read op arg 3; if (int)key < 0 → no colorkey (opaque); else the operand is 0xRRGGBB, converted to 0xFFRRGGBB and passed to the surface creator FUN_00477c40 as the transparent key (so operand 0 = key black). Colorkey is baked at surface-load (matching texels → transparent), NOT compared per-blit. Port: interpret arg 3 as RGB888; <0 = none; else texels whose (R,G,B) equal the key become transparent when the surface image is loaded/cached.
• 0x202 (gfx_op_0x202_worker_set_color_anim 0x47ea00): sets an animated color/alpha target obj+0x64 = packedARGB, the color-anim active bit, resets progress obj+0x34=0. Animates over the global clock (0x238).
• 0x203 (gfx_op_0x203_worker_set_color 0x47e9b0): sets a static color/alpha obj+0x60, no anim bit. Immediate per-object modulation.
• Blit (gfx_object_blit_d3d9 0x4774c0): selects a blend mode (local_2c: 0 opaque, 1 alpha SRCALPHA/INVSRCALPHA, 2/3 additive/special for glow/flash) and passes a modulation color/alpha to the device draw. Slice A ports the alpha path (fades); additive (glow) is deferred (its local_2c source field is not yet pinned).

Interpolation RE pass (2026-07-08, stalled → both deferrals confirmed). Attempted to pin how a 0x202 fade animates so smooth ramping could join slice A. Findings (Ghidra gfx_object_anim_interpolate 0x473ed0, annotated+saved): the bit-2 anim family (op 0x234) interpolates 5 independent sub-channels (color obj+0x240/period obj+0x220, two matrices, rotation 0x168=360°, src-rect scroll), each on the global frame clock ctx+0xb550 (advanced per present, NOT the op-0x238 clock ctx+0x51b7c), and each ping-pongs (triangle wave, folded at period/2) — i.e. these are oscillating/pulsing effects, not one-shot fades. The 0x202/0x203 color (obj+0x60 static / obj+0x64 animated, sets ctx+0xb558/0xb560) is a separate channel whose blit consumer was not located in this pass. So a one-shot fade's exact source→target→easing is still unresolved and would take a dedicated dig (find the obj+0x64 consumer + the color→obj+0x240 path + the clock advance). ⇒ smooth color-anim interpolation stays deferred; slice A ships the static end-state (which reaches the correct final alpha/tint and fixes the stuck-opaque bug), with interpolation as a scoped follow-up.

SC0000 anim/transform/spritesheet cluster — op→field map (2026-07-08)

Reversed for the animation cluster slice (spec docs/superpowers/specs/2026-07-08-sc0000-anim-transform-cluster-design.md). Every cluster handler resolved via the dispatch table handler(op)=ctx[0x26c93+op] (the opcodes.toml u004xxxx labels are Kelebek VA drift — do not use them). Each op is a thin wrapper (FUN_0041b940(n) fetches operand n) → a worker that writes object fields; the interpolator gfx_object_anim_interpolate (0x473ed0) is the consumer. The cluster is heterogeneous — setters, queries, and a movie op. Renamed + annotated in Ghidra, saved.

In scope (built this slice):

op	handler / worker	semantics
0x22f	gfx_op_0x22f_set_position_anim → gfx_worker_set_translation	set object position (translation vec obj+0x5d4); base transform, not a ping-pong channel
0x229	gfx_op_0x229_set_position (FUN_00472bb0+FUN_00472be0)	set object position/geometry immediately (obj+0x420/0x424 + vec obj+0x440..0x448)
0x239	gfx_op_0x239_set_srcrect_cell → gfx_worker_set_srcrect_cell	set spritesheet grid obj+0x238/0x23c + static cell obj+0x234
0x231	gfx_op_0x231_anim_srcrect → gfx_worker_anim_srcrect	animate spritesheet cell: bit2 active, period obj+0x230, grid obj+0x238/0x23c → interpolator SRC-RECT SCROLL channel (ping-pong across the row)
0x232	gfx_op_0x232_anim_color → gfx_worker_anim_color	animate color (pulsing GLOW): bit2 active, period obj+0x220, target obj+0x240 → interpolator COLOR channel (ping-pong). Distinct from static 0x202/0x203 (obj+0x60/0x64)
0x228	gfx_op_0x228_query_position (FUN_0047cdd0)	query current computed (x,y,z) → operand slots 3/4/5 (script logic, not render)
0x23f	gfx_op_0x23f_query_object (FUN_0042a520)	query an object status/value → operand slot 1

Deferred (own follow-ups, per scope decision): 0x21f (FUN_0047eb70, 4-float scale/matrix), 0x223 (FUN_0047f440, 8-arg matrix row) → need affine rendering; 0x236 (gfx_op_0x236 @0x423ee0) a timed/animated-surface (movie-like) op; plus the unclassified 0x21c/0x21d/0x224/0x242/0x23d/0x20a/0x20e/0x243 tail (2-arg flags / inline). These stay GAP until a follow-up slice or are safe-noop'd if the opening tolerates it.

Grey-background root cause — slot collision + tint-strength (2026-07-08, gfx-log)

Diagnosed with the new --gfx-log compositor/op trace (docs/tools-reference.md). The grey background has two distinct causes, both now proven:

1. CORRECTION to the blend section above — op 0x202/0x203 "alpha" is a TINT STRENGTH, not object opacity. Evidence: the primary CG is drawn with 0x203 (alpha=0, color=white) = 0x00ffffff. That means "blend the tint (white) into the texel by strength 0" = no tint, fully opaque CG — but slice-A treated the alpha byte as the object's opacity → the CG rendered fully transparent → grey. Fix (commit 5e4fdda): RenderObject.TintStrength split from Alpha (opacity); textured objects stay opaque and the tint LERPs the RGB by strength (0=keep texel, 1=full tint). Surfaceless fills use the strength as fill opacity. Verified: the opening event CGs render again (shot-confirmed).

2. Effect pages: everything collapses into slot 0. set-texture is dominantly set-texture (GLOBAL resId)(GLOBAL slot)(local colorkey) (543× across the corpus); the slot is a global. In our run every such global resolves to 0, so the background (BG030A), event CGs, and the effect spritesheet (AE001H, an 800×400 4×2 grid of blob frames) all set-texture into slot 0. Objects live-reference their slot, so loading the effect evicts the BG → grey; and the effect is drawn full-screen from slot 0 (its object src=(0,0 800x600)) → the whole sheet (blob grid) covers the screen. ⇒ The layering failure is a slot-assignment problem. CONFIRMED CAUSE (2026-07-08, live AGE_DIAG_SETTEX trace): every set-texture slot = G[0x62452], written by query-gfx-object? (0x215) which returns -1 for the (correctly-unregistered) CG/effect handles → the fallback at SC0000 label_12649 does G[0x62452] = lookup-array-2d(rec[s3]=G[0x3239], G[0x62450], 3, 0) = 0 because the slot table rec[s3]/G[0x3239] is empty. That table is filled by call label_125bd (SC0000 0x50f, slots 4..13), which is reached only through the scene-coroutine framework — the G[0xaba5c] gate (0x450) + op 0x140 (u0041F9C0, coroutine LABEL/yield "LABEL" "J" @ 0x46d). Op 0x140 is stubbed (SC0000 GAP list) → the coroutine re-entry never routes through label_125bd → slot table stays 0 → all layers collapse into slot 0. This is exactly the "second still-latent gap" flagged above (§"the render drift's second half"), now confirmed as the cause of the visible grey BG on multi-object pages. Fix = implement the scene-coroutine framework (0x7b/0x7c/0x140 + G[0xaba5c] gate) so label_125bd runs (or, as a targeted unblock, run label_125bd/seed G[0x3239] directly). NOT a compositor/z-order/blend bug. Diagnostics: AGE_DIAG_SETTEX=1 env → VM logs each set-texture slot operand + query-gfx-object? result.

---

Native walls backlog (targets for this loop)

• call-script dispatch — SOLVED (above): call-script <id> = raw SYS4INI file index.
• decision→scene — how 0x62ccf/the decision selects the next SCxxxx. Now narrower: scenes load via call-script/the same SYS4INI-index loader, so the open question is only where the decision value is turned into a scene id (a caller of SCJUMP; re-aimed away from u00428010).
• op 0x60 (u0041A270) — the rand-like value gating 1732/1755 SCJUMP decisions.
• gfx command-buffer — DONE (the 0x212–0x21a positioned-object subsystem = the rendering drift): all 14 ops reversed + implemented against a host-side GfxState, and the missing INIT2 boot state supplied via --boot. CGs render (screenshot-confirmed). See the op 0x215 finding + "The render drift's SECOND half" above. Remaining: AE* alpha/blend (deferred) and cold-object anchors.