Strict lockstep over-reported: the operand hook misses any op whose operands
aren't fetched via vm_operand_fetch (comment 0x1a7, set-string 0x192, the
0x1c7/0x1cc/0x131/... string-op cluster) — the engine executes them (they sit
after non-branching ops) but they're absent from its trace. New align() resyncs
over such one-sided insertions and reports only NON-realignable forks; it
separates VM-only blind spots (artifacts) from engine-only detours (real,
reconverging branch/state gaps, surfaced honestly). +2 tests (7/7).
Result: with G[0x6c1] seeded the SC0000 opening has NO non-realignable fork
across all 539 VM ops (was: false 'diverge at 0x8d'); residual = a 2-op
engine-only color detour (0x202/0x203 @ 0x122d0). Cold's first real fork is a
later G[0x6c1] gate. Validates the pre-scene-state theory end-to-end.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Operand-hook captures skip zero-operand ops (stmt markers, script-entry 0x259),
so the VM offset trace is filtered to argc>=1 instructions before diffing — same
subsequence both sides. +1 unit test. Default on; --full for a tick-mode capture.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Reads cur_ctx_index/frame_pc/frame_codebase from the engine ctx per executed op
and emits (codebase, offset=(pc-codebase)/4) to build/engine-optrace.jsonl.
Recon result: the tick hook (0x410fb0) does NOT expose ctx via ecx (0 entries);
the operand hook (0x41b940) is the working capture (100% of offsets land on valid
SC0000 instruction starts). Writes a tracer-live.flag so the capture can be gated
on hooks-installed before the scene loads (else the entry burst is missed).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Pure first_divergence + pick_scene_codebase (longest-common-prefix codebase
identification), unit-tested (test_diff_optrace.py, 4/4). CLI loads the engine
jsonl + VM json, isolates the scene's codebase, and reports the first divergence
with the mis-modeled instruction and +/-3 ops of context on each side (opcode
+ rendered line via sys4load).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Canonical source vm-map/engine-ctx.toml (35 documented ctx fields) +
engine_ctx_build.py (--build/--lint, unit-tested) -> build/engine-ctx.json
+ docs/engine-ctx-reference.md. Applied to Ghidra in the next task.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
248 module-resident core imports at the RVA 0x16f000 rebuilt IAT
(kernel32/user32/gdi32/winmm/advapi32/ole/oleaut/version/ntdll);
29 singletons set aside. Anchors confirmed: ReadFile/CreateFileA/
SetFilePointer + timeGetTime@0x16f3d4 (=DAT_0056f3d4). d3d9/shell32
etc. are heap-resolved (out of dump) — expected.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
GATE-PASS: 23342 exports mapped, 277 in-range pointer matches clustering
into a module-resident import table at RVA 0x16f000 (VA 0x56f000);
23 singletons. ~254 real imports vs pe-sieve's 17-in-noise.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Extract op->real-handler map (handler(op)=ctx[0x26c93+op]) from the
registration routine's override stores; ghidra_handler_map.py +
build/op-handler-map.json (420 overrides). Cross-check vs opcodes.toml
found 0 real drift. One-shot Ghidra pass then labeled the /v2 image:
281 raw FUN_/LAB_ handlers -> op_0xNN_handler, 107 bare VAs -> functions,
31 hand-named preserved, opcode plate comment on every handler.
Includes the Task A spec + plan and the two-program (/v2 vs SMM) gotcha.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Design artifacts for the merged frame-stepped VM work (throttle the Godot VM
to a per-frame op budget). Probes measured the native ~1788 ops/sec cadence
and uncapped D3D9 Present that motivated the wall-clock-op-rate approach.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add tools/scene_opcode_coverage.py: histograms a scene's static opcodes and
classifies each vs the C# VM as impl / safe-noop / GAP (effectful op the VM
silently stubs). Implemented set is parsed live from VirtualMachine.cs case arms
(no drift); metadata from build/opcodes.json. Makes a half-rendered scene legible
("N ops still stubbed") instead of implying everything runs.
SC0000 baseline: 129 distinct ops, ~94.8% instruction-weighted handled, 68 GAP.
The tracker cross-checks opcodes.toml vs VM behavior and surfaced 0x259
(script-entry marker) missing its noop_headless flag -> reconciled in opcodes.toml
and rebuilt (regen: age_opcodes_himegari.py, opcode-reference.md).
Docs: tools-reference.md (tool row), phase-a-slice-plan.md (completeness gauge).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
VM dispatches on the opcodes.toml label field (OpcodeTableJson), so the 14 gfx
ops' labels are set to their dispatch strings. Ops still stubbed until the VM
cases land (Task 3.3+); full suite green (34), lint clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Ghidra-read all 13 still-stubbed gfx-family handlers via the dispatch table
(0x1a2/0x1f7/0x1fa/0x1ff/0x202/0x203/0x212/0x213/0x216/0x217/0x218/0x219/0x21a);
recovered 0x219 (was unanalyzed). Every op writes a cmd-type into the current
object record then SETs fields (native worker) or QUERYs fields (returns to
operands). Recorded the op-contract table in engine-re.md, set opcodes.toml
semantics (source=investigation, confidence=high), renamed handlers in Ghidra.
Key finding: the drift has TWO stubbed drivers, not one — 0x215 (slot-select)
AND 0x218/0x21a (per-object geometry vectors feeding the anchor-preserve math).
The subsystem is a per-object record model (slot + position/anchor 3-vectors +
color) queried by the ops; native DirectDraw workers need not be modelled. This
is the spec for Phase 2 (host-side model design).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Read op 0x215's real handler FUN_0042a0b0 (resolved via the dispatch table
ctx[0x26c93+op]; Kelebek's 0x421160 is VA-drift). It writes cmd-type 5 into the
current gfx-object record and returns a std::map::find over an engine-internal
registry populated by sibling gfx ops (0x1a2 hash insert). The return is native
command-buffer state, not the VM global bank -> seeding story-state cannot fix
the drift. Verdict: (b) a genuine native op, NOT (a) state-divergence.
Reconcile the previously contradictory drift accounts onto one canonical home
(engine-re.md op 0x215), with opcodes.toml carrying the opcode-level semantics
and phase-a-slice-plan / tools-reference / frida README corrected to point at it
instead of repeating the disproven state-divergence conclusion.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Native-RE (Ghidra) cracked call-script <id> (opcode 0x03): its handler
FUN_0041bc90 -> loader FUN_0040e980 -> resolver FUN_0044f390 indexes an
80-byte record table at base + id*0x50 == the SYS4INI record layout. So
`call-script <id>` is a direct RAW index into the SYS4INI global file
table (the asset index we already parse) -- there is no separate on-disk
id->code registry. This resolves name-resolution.md #1, statically, no
Frida.
Confirmed: all 297 distinct corpus call-script ids resolve to a .BIN
script with a semantically-exact name (0x1ab->ADDITEM, 0x2ae7->MES,
0x143->BUNKI, 0x329d->CALCREVISE), 0 out-of-range, 0 alternate-pack.
Companion op 0x8f `call` is an intra-script JSR (FUN_0041fba0), not
cross-script.
- parse_sys4ini.py: preserve `raw_index` per entry (= the engine file id;
index the RAW records incl. '@' placeholders) + emit
build/callscript-names.json (id->name).
- sys4load.py: annotate `call-script 0x1ab =ADDITEM.BIN`.
- opcodes.toml 0x03/0x8f refined (source=investigation, confidence high,
handler VAs) + rebuilt opcode-reference.md.
- docs: engine-re.md (op 0x03 section + backlog re-aimed),
name-resolution.md #1 (SOLVED), script-inventory.md (call graph +
living-reference decision), tools-reference.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
op 0x1a2 resolve-handle? -> gfx-cmd-register (verified handler FUN_0042d360, gfx
command-buffer op). name-resolution: decision->scene hop is native+unidentified, not
u00428010 (disproven). Regenerated opcode-reference + shim.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Reworked from static-witness-only to execution-driven: ~99% of SCJUMP decisions
are gated by a native computed value (op 0x60), so witness-synthesis alone can't
cover them. Value-local tracking resolves load-then-compare guards to real globals;
native terms honestly marked opaque. Verify: static 3/3 exact + 279/279 execution-
driven consistent over 2000 seeds, 0 failures.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
build/globals.json is generated (build/ is gitignored, regenerable via
globals_build.py --build); docs/global-reference.md is the tracked human view.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Live capture (esi=engine ctx via operand-fetch, poll object-record array
[esi+0x53d64] stride 120). KEY FINDING: through the full real opening, the record
array holds only 3 persistent UI objects — NO CG objects. The real game does NOT
draw opening CGs via the 0x212-0x21a positioned-object path our headless VM uses;
with state it takes a different (direct) branch. So the bg/sprite drift is a
STATE-DIVERGENCE artifact of the unseeded headless VM, not a missing native op —
the fix is the Phase B state/choices flow (makes label_12649 take the if-branch).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
operand-fetch (call 0x41b940) fires ~8500/sec => the VM interpreter executes
from the in-place unpacked module at 0x400000 (NOT the heap copy) => handlers are
hookable by dump address. gfx-family(0x212-0x215)=0 at the title (no CG commands
until a scene runs).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Dump validated: AGE.EXE is unpacked in-place at 0x400000 in memory (AGF-decoder
landmark @0x474f1f reads real code with the 'BM' 0x4D42 check). Kelebek handler
VAs map directly (VA-0x400000 = file offset). Handler ABI: thiscall (esi=engine
context), operands fetched via call 0x41b940, per-object command-type table at
[esi+idx*120+0x53d88]. 0x215 is part of a native gfx command-buffer manager.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Both on-disk engine images are the same packed binary, so native handler code
exists only unpacked in memory. dump_engine attaches, enumerates ranges, and
dumps the AGE.EXE module + large r-x heap regions (chunked) to build/engine-dump/
for offline disassembly (locate 0x215 @ VA 0x421160 via the dispatch table).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Investigation upgrade: 0x215 queries the native object manager by element
handle-id (0x62455[idx]) and returns the object's slot/status; sign-tested to
drive label_12649's slot-select. Stubbing it collapses all draws onto slot 0 →
the anchor-preserve geometry reads foreign textures → cumulative bg/sprite
drift. Keystone for the graphics object-manager; exact return via unicorn
(handler @0x421160).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
C# VM is byte-identical to vm0.py across all 297 SC/SP scenes (offsets+halt+steps);
RECOVER passes; SC0000 = 27994 steps / 186 lines matching the Python prototype.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Reverse-engineering + open reimplementation workspace for Eushully's AGE/SYS4
engine (first target: Himegari). The repo root is age-reimpl/; the original game
install and the extracted ALF data are siblings outside the repo and are never
tracked. build/ (derived corpora) is gitignored and regenerated by the tools.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
