Files

gamer147 0ae4f607c0 docs: native-engine RE — Ghidra+MCP loop + opcode dispatch table + u00428010 correction

Ghidra+MCP workflow validated. Found the opcode->handler dispatch table
(*(ctx+0x9b8f4+op*4), registered by FUN_00413860) — the general fix for Kelebek VA
drift. Corrected: u00428010 (op 0x1a2) is a save/resource op, NOT decision->scene;
the SCJUMP consumer persists the visited-decision flag. New doc docs/engine-re.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

2026-07-07 11:29:02 -04:00

3.6 KiB

Raw Blame History

SCJUMP — progression / scene-sequencing logic

SCJUMP.BIN (29,790 instructions) is the game's progression state machine: it decides "what comes next" in the story. tools/scjump_decode.py decodes its decision logic faithfully into a readable, VM-verified table.

Mechanism

Top switch on 0x3234 (chapter_mode, 1..9) dispatches to a per-chapter block (1→0x81, 2→0x9f, 3→0x1ded, 4→0x1e6d, 5→0x54c1, 6→0x93c4, 7→0x1e868, 8→0x2aa42, 9→0x2b8a3).
Each chapter block is an acyclic tree (0 back-edges) of comparisons on progress counters (scjump_progress_a/b/c/d = 0x4dfbc/0x2052e/0x152618/0xe6c5d), story flags, and a native computed value (see below).
At each of 1755 decision sites it writes 0x0 = 1 (decision-valid) and 0x62ccf = <decision value> (847 distinct values, 0..1693).

The decode

scjump_decode.py walks the CFG with a guarded DFS (a guard stack + a local→symbolic-value map, so mov local, global ; cmp local, imm resolves back to a real global condition) and emits, per site, (chapter, path-condition guards) → decision. Output (generated, disposable): build/scjump-decisions.json (machine) and build/scjump-decisions.md (human, globals rendered by their vm-map/globals.toml registry names). Regenerate: py -3.11 -X utf8 tools/scjump_decode.py.

Native-gated decisions (a key finding)

1732 of 1755 decisions have at least one guard on a native computed value — op 0x60 (u0041A270(local, 0x64), e.g. local = f(100), then local < 25). This is resolved at runtime by the engine (it looks like a random/derived roll: ≈ rand()%100 < N), so it is not recoverable statically and is honestly marked opaque in the guards. The decode still captures every global condition (chapter, counters, story flags) on each path; only the native term is opaque. Takeaway: SCJUMP mixes deterministic story-state gating with native/random gating — many "what comes next" choices are not pure story flags.

Verification

--verify cross-checks the decode against the executing VM (vm0.py) two ways:

static-witness — for the fully-static decisions (no native/opaque guard), synthesize a state from the guards and confirm the VM emits that decision (3/3 exact).
execution-driven — seed thousands of random states, run SCJUMP (the VM executes the native op too), and for every realized decision assert the table has that site with the right value and no static guard is violated by the realizing state (279/279 consistent over 2000 seeds, 0 failures). This is the oracle that covers native-gated paths. Run: py -3.11 -X utf8 tools/scjump_decode.py --verify.

The native decision→scene boundary (deferred)

Consumers do lookup-array(ptr, 0x5f0ed, 0x62ccf) then u00428010(ptr). Correction (2026-07-07, via Ghidra): u00428010 (op 0x1a2) is not the scene resolver — it's a save/resource-file op (its handler formats %s\SAVE%2.2d.DAT), so that pattern persists the "visited-decision" flag into the save, not loads a scene. See docs/engine-re.md. The actual decision→scene resolution is a different, still-native mechanism — most likely call-script/script-load-adjacent (name-resolution.md §1). Cracking it uses the engine-dump + Ghidra loop (the opcode-dispatch table found in engine-re.md is the key); still a separate slice.

3.6 KiB Raw Blame History