OpenMaidEngine/docs/phase-a-slice-plan.md

Phase A — Vertical Slice Plan (the first build step)

Concrete execution plan for Phase A of remake-architecture-and-roadmap.md. Decided over the alternative (fully decoding SCJUMP.BIN) after recon showed SCJUMP is not the gating unknown.

Why the slice, and why headless-first

SCJUMP recon (2026-07-06): SCJUMP.BIN is a 29,796-instruction progression state machine, not the call-script registry. Top level switches on global 0x3234 (mode 1–9 → big blocks); each block is nested eq/ne/and/jcc on flags, ending in movs to output globals. Almost no call-script. So it decides what scene/branch comes next via state, and does not resolve call-script id → code. Consequence: the id→code registry stays engine-level (deferred), but the slice can stub call-script — it is not gating for running one scene's dialogue.

Correctness bootstrap (roadmap §5) drives the ordering: the VM must be validated-correct before it is trustworthy. Our strongest oracle is build/text/dialogue.jsonl (the show-text lines per script). So the very first slice is headless and text-only, validated by that oracle — no Godot, no AGF, no audio, no dispatch registry. Only once the VM reproduces dialogue do we add rendering.

Phase A therefore splits:

• A0 — headless VM, dialogue-validated (Python prototype). ← immediate, executable now.
• A1 — port the validated model to C# (the runtime's VM core).
• A2 — Godot ADV backend (render one scene with visuals + voice).

---

A0 — Headless VM validated by the dialogue oracle

Goal: a Python interpreter that executes one ADV scene's bytecode and emits its show-text sequence; that sequence is a coherent, in-order subsequence of the script's static dialogue.jsonl lines. This proves the execution model — control flow, operand/pointer semantics, string handling, and the no-op-marker assumptions — before any C#/Godot investment. Reuses tools/sys4load.py for all parsing/decoding (no new parser).

Execution model to implement

• Memory: one flat global bank = dict[int,int] (globals are raw offsets into one space; global-int A ⇒ G[A], default 0). Per-call local frame with typed banks sized by header F0–F5 (local_int[F0], local_float[F1], local_string[F2], …).
• PC / control flow: build offset→instruction-index map from sys4load instructions (each has .offset = dword index; jump targets are dword indices). jmp t → pc = map[t]. jcc(cond, A, B) → cond truthy ? goto A : goto B, where 0xffffffff = fall through (confirmed model from RECOVER).
• Operand resolution by type: imm→value; global-int→G[value]; local-int→frame.int[value]; string(2)→decoded string at dword offset; float/global-string/etc. analogous.
• ⚠ Pointer/lvalue semantics — the key modeling task. RECOVER proves -ptr operands are lvalues: lookup-array(dst_ptr, base, idx) yields a reference to G[base+idx]; mov through a ptr writes to the referenced cell; reading a ptr rvalue dereferences it. Model a ptr slot as holding an address into the global bank; nail this so the RECOVER array-copy produces correct results (unit test it directly).
• Opcode handlers (~52 named ops):
  • arithmetic/bit add sub mul div mod and or sar shl → p1 = p2 ⊙ p3.
  • compares eq ne lt lte gr gre → 0/1.
  • mov (incl. through ptr), lookup-array (p1=mem[base+idx]), lookup-array-2d (p1=mem[base + i*stride + col]), copy-to-global, set-array-to, bit-set/reset, check-bit.
  • control jmp call jcc ret exit exit-script.
  • string set-string concat strlen toString.
  • ADV capture: show-text → append (arg text) to the emitted list; end-text-line, wait-for-input, set-font, comment → capture/skip (no visible state).
• Markers → no-op (this TESTS the classification): 0x1f4 0x1f5 0x1d5 0x1bc 0x1bf skip; tentative 0x21b 0x1d2 0x258 skip — if dialogue stays correct, the no-op assumption is validated.
• call-script → STUB: log (id), return immediately. (Its dialogue belongs to other scripts; stubbing keeps the emitted set = this script's own lines.)
• Effectful (draw/texture/audio/ui/input) → STUB: log and ignore.
• Unknown/other opcodes → log + no-op, so a rare op doesn't halt the run (record coverage).

Oracle & scene choice

• Oracle: with calls stubbed and default state, every emitted show-text line must be a real decoded string from the script's pool, and the sequence must be an in-order subsequence of that script's dialogue.jsonl lines (≈ equality for a linear scene). Catches: garbage strings (bad operand/ptr handling), impossible ordering (bad control flow), missing/extra lines.
• Scene pick: choose a short, mostly-linear ADV scene — high show-text count, low jcc density, few call-script. Selection step: rank SC####/SP#### by (show-text count) / (jcc + call-script count), small size. Known-good fallback: SC0030.BIN (dialogue verified). Also run a RECOVER unit test to validate pointer/array semantics independent of dialogue.

Steps

1. tools/vm0.py: load a script via sys4load, build offset→index map, frame + global bank.
2. Implement operand resolution + the arithmetic/compare/mov/lookup/control handlers; unit-test on RECOVER.BIN (array copy + both loops must produce correct global writes).
3. Add ADV capture + markers-as-noop + call/effectful stubs; add opcode-coverage logging.
4. Run on the chosen linear scene; diff emitted show-text vs dialogue.jsonl (subsequence check); eyeball the first ~15 lines for coherence.
5. Iterate until several scenes pass; record which ops/markers were exercised and any surprises (esp. whether the tentative-no-op markers hold).

Success criteria (A0 done)

• RECOVER unit test passes (pointer/array model correct).
• ≥3 ADV scenes: emitted show-text is a coherent in-order subsequence of their dialogue.jsonl, no garbage strings.
• Coverage report of which opcodes actually executed (drives A1/A2 priorities).
• The no-op-marker assumption is confirmed or corrected with evidence.

---

A0 result (2026-07-06) — execution model VALIDATED

tools/vm0.py built (reuses sys4load; ~250 lines). Results:

• RECOVER unit test PASSES — all 7 checks (block-1 3-field copy, block-2 restore + flag, both skip-guards). The pointer/lvalue model, 2D stride indexing, both loops, and two-way jcc all execute correctly. The core execution model is proven.
• Full SC/SP oracle sweep (vm0.py --sweep): 282 / 294 scenes DIALOGUE-VALID = 95.9%. Every emitted show-text line is checked (by string offset) as an in-order subsequence of the script's static dialogue.jsonl lines. Zero STRAY and zero ORDER violations across all 294 scenes — the model never emits a garbage string and never emits dialogue out of order. 279 CLEAN (valid + natural exit); 3 OK/LOOP (valid subsequence, halted by the loop-guard); 12 EMPTY; 3 skipped (no static show-text). SC0000 = 326 static / clean; SP0062 = 220/220 CLEAN.

A0-remainder work done (2026-07-06, session 2):

• Loop-guard added (EMIT_CAP=2): halt a run once any single line is re-emitted a 3rd time — a semantic guard tied to the oracle (vs. a blind step limit), and it classifies the scene LOOPED instead of spewing garbage. The 3 zero-state spinners (SC0010/SC0600/SC0200) now terminate cleanly in <12k steps and their emitted lines are all valid.
• SP0062 "stray" was a measurement artifact, not a bug — the precise offset-based oracle shows it CLEAN (220/220, natural exit). Offset-match ⟹ text-match (VM decodes each string at the same offset the extractor did), so CLEAN is trustworthy.
• 0x71 (label-def) folded into the no-op marker set — structural, no runtime effect.
• Sweep + single-scene diff harness added to vm0.py: --sweep [N] (coverage table over all SC/SP), --scene NAME (detailed diff for one script), plus load_oracle/subsequence_status.

op 0x90 investigated in depth — it is input chrome, NOT a correctness hole (full evidence: vm-map/opcodes.toml op 0x90 details). Kelebek left it "ukn"; corpus analysis resolves it: 0x90 x y w h tgt_a tgt_b tgt_c (argc 7) is a cursor/input hotspot hit-test that branches per interaction outcome and falls through to pc+1 when nothing matches (design-confirmed: enc.len 15 lands the next instr on the fall-through statement). It occurs ONLY in a shared ADV-chrome subroutine that is byte-identical in all 301 ADV scripts — exactly 8 sites each (5 immediate-rect buttons at (684..772, 572) toggling G[0x6c9..0x6cd] + 3 local-operand keyed forms), zero scene-specific use. Headless (no cursor/input) ⇒ fall through ⇒ vm0's stub is already correct, proven safe by all 279 CLEAN scenes (which contain these same 8 sites). op 0x97 (argc 5, no targets) is its companion register-hotspot call. So 0x90 stays as fall-through in A1 with confidence; it is modelled as a live hotspot test only in A2 (Godot input backend), confirming target→state mapping via Frida.

The 12 EMPTY scenes — state-gated interactive screens, not a model failure. Traced SC0830: it exits early because G[0xaba5c]==1 gates the content; past that gate the dialogue sits behind the ADV input-wait loop (the hotspot-polling chrome above), so with no seeded state and no input the scene exits or spins before reaching text. Unlocking them = seed per-scene state + supply input → Phase A2/B, not an A0 model fix.

⚠ Honest scope of the 95.9%: the subsequence oracle proves no-garbage / in-order, not a complete path — inherent to a subsequence oracle run headlessly (interactive/state-gated branches take the no-input path by design). That anti-garbage guarantee is exactly what A0 set out to prove.

Confirmed by this run: the classified no-op markers (0x1f4/0x1f5/0x1d5/0x1bc/0x1bf + tentative 0x21b/0x1d2/0x258, now + 0x71) are safe as no-ops for ADV flow; call-script is stubbable; effectful ops (draw-texture/create-texture/play-voice/0x1f7/0x202/0x203/…) stub cleanly.

✅ A0 COMPLETE. Success criteria met: RECOVER unit test green (pointer/array/control-flow model proven); 282 ADV scenes emit clean in-order subsequences with zero garbage; coverage number recorded; no-op-marker assumption confirmed at scale; op 0x90 (the last big control-flow unknown) resolved as input chrome whose fall-through stub is correct headless. Next = A1 — port the model to the C# VM core, differential-test against vm0.py. 0x90/0x97 stay stubbed (correct headless); the interactive input path + per-scene state seeding land in A2 (Godot backend) alongside the real hotspot model.

A1 — Port the validated model to C#

Reimplement the A0 execution model as the runtime VM core in C# (the language decision from the roadmap; GDScript is too slow for the loop). A0 is the reference: differential-test C# against the Python prototype's traces on the same scenes. Port the container parser too (or load via a shared spec). Deliverable: headless C# VM reproducing A0's results.

A2 — Godot ADV backend (one scene, with visuals)

Wire the C# VM's effectful ops to Godot: show-text/message window (+ furigana via display-furigana), set-font, wait-for-input, choices, play-voice/play-bgm, and create-texture/set-texture/ draw-texture/draw-string for the background + sprites. Convert the scene's AGF art with the on-disk AGF2BMP2AGF.exe. Resolve just-enough call-script/state so the scene's setup runs (or hand-set the preconditions). Deliverable: the chosen scene playable in Godot — bg + dialogue + a choice + voice — matching A0's text.

---

Risks / open questions for A0

• Pointer/lvalue semantics — the main modeling risk; RECOVER is the litmus test.
• Initial global state — a scene may assume preconditions from earlier flow (SCJUMP/prior scenes). Mitigation: default-zero globals + set the few a scene reads early; the subsequence oracle tolerates a shortened path.
• Runtime vs static dialogue order — static dialogue.jsonl is file-order (all lines); runtime is execution-order (branch taken). Hence subsequence, not equality; pick linear scenes to tighten it.
• Hidden effect in a "stub" — a stubbed effectful op that actually gates control flow could skew output. Watch for divergence; promote a stub to a real handler if a scene needs it.

Immediate next action

Build tools/vm0.py and get the RECOVER unit test green (pointer/array/control-flow correctness), then run the first linear ADV scene against the dialogue oracle. That single result tells us whether the whole VM approach executes correctly — the load-bearing question behind option 3.