OpenMaidEngine/docs/phase-b-framework.md

Phase B Framework — Natural Boot to First Gameplay

Phase B broadens the proven ADV vertical slice into a naturally booted, stateful play session and then into the first narrow gameplay loop. This document is a sequencing framework, not a task-level implementation plan. Phase A remains active until SC0000 meets its completion criteria.

The architectural preference is:

complete SC0000
  -> persistent session and scene coordinator
  -> faithful system/data boot
  -> title and New Game happy path
  -> SC0000 under naturally initialized state
  -> natural first-dungeon transition
  -> bounded first-dungeon gameplay slice

This order makes gameplay failures attributable to gameplay rather than to missing boot state, discarded globals, or manually inherited host state.

Entry criteria from Phase A

Phase B may begin when SC0000 is a reliable presentation baseline:

• A normal windowed playthrough is visually and audibly coherent from entry to natural exit.
• Reproducible presentation inconsistencies have been resolved or explicitly classified with evidence.
• Every SC0000-executed effectful opcode is implemented, or its lack of a host-visible effect is supported by native/script evidence. Remaining unrelated corpus gaps do not block entry.
• The run no longer relies on unexplained timing, input, layer, or resource workarounds.
• Automated VM/host regressions and a repeatable manual playthrough form the acceptance baseline.
• SC0000's terminal global state and control-flow boundary can be captured for comparison once natural scene chaining exists.

The Phase A implementation/result history remains in docs/phase-a-slice-plan.md.

Principles

1. State correctness before systems breadth. Dungeon logic depends on initialized unit, item, skill, progression, configuration, and heroine state. Establish their natural producers before debugging their consumers.
2. Follow script control flow. The bytecode owns game rules. Implement the effectful operations and lifecycle services it calls; do not replace dungeon/combat logic with a parallel rules engine.
3. One natural path first. Boot → title → New Game → SC0000 → first dungeon is the initial spine. Alternate menu branches and broad gameplay coverage grow from it later.
4. Persistent session, replaceable scenes. Globals and game/profile state survive scene changes while script frames, retained presentation state, and scene-owned resources observe proven lifecycle rules.
5. Demand-driven opcode work. Investigate an unknown opcode when the chosen path executes it or evidence connects it to a reproduced defect.
6. Bound every slice by an observable transition. Each stage starts from a known state and ends at a visible screen, input boundary, scene handoff, or gameplay action.

Stage B0 — Ground-truth reconnaissance

Before changing runtime architecture, record the original game's path from process start through the first meaningful dungeon interaction. The goal is an answer key, not exhaustive reverse engineering.

Capture:

• Script/load order across system boot, title, New Game, SC0000, and first dungeon entry.
• Which initialization scripts run and which global banks or native/profile values they establish.
• Retained graphics/audio state that survives each boundary.
• The title selection and New Game dispatch path.
• SC0000's natural terminal decision and the corresponding next loaded script.
• The first dungeon's executed opcode/call-script families, assets, and obvious state dependencies.

Detailed progression semantics remain canonical in docs/scjump-progression.md; native loader findings belong in docs/engine-re.md and docs/name-resolution.md.

Initial B0 result (2026-07-20)

Static SYSTEM4/INIT2 control flow plus an existing native opcode trace establishes the first natural spine:

SYSTEM4 → config load/init → INIT2 (+23 nested data initializers, then TUNE) → optional LOGO/OP → INIT → TITLE → GAMESTART → UNITECH/CALCARR → TUNE → TITLE return → SYSTEM4 → SC0000.

SYSTEM4, not an opaque native dispatcher, is the long-lived scene coordinator. It maps the SCJUMP decision through a global resource-id table, places the result in G[0x699], and uses computed call-script; the initial zero decision falls back to raw SYS4INI id 0x22, SC0000.BIN. The current VM already supports computed nested call-script frames. Consequently B1 should preserve one VM and host rooted at SYSTEM4, letting script-owned setup/cleanup surround child scenes, rather than invent an out-of-band replacement protocol. Full process-start observation remains useful for profile/default and retained host-state evidence, but is no longer needed to guess the script coordinator architecture.

The current headless C# runner already follows this root naturally: one SYSTEM4 run entered INITCONFIG, INIT2 and all 23 of its data-initializer children, TUNE, INIT, and TITLE (28 nested script calls total), then remained in TITLE's input-poll loop because the diagnostic host supplies no user input. Direct opcode coverage is 100% for all 23 data initializers, CALCARR, and TUNE; the remaining direct coverage is SYSTEM4 64/82, INIT2 9/12, TITLE 61/65, GAMESTART 43/47, and UNITECH 29/31. B0/B1 should therefore make the SYSTEM4-rooted path visible and interactive in Godot, then investigate only the gaps actually reached on that route instead of treating every static gap as a prerequisite.

Godot root landing (2026-07-20). The no-argument Godot/run-godot path now starts SYSTEM4 directly and does not apply the direct-SC0000 layout/surface bootstrap or the diagnostic --boot prefix. A windowed run reaches and renders TITLE using SYSTEM4-owned retained state. A real-script integration test drives TITLE's Game Start input, GAMESTART's release-gated default selection, and proves the same VM enters SC0000 through SYSTEM4's computed resource id G[0x699]=0x22; G[0]=1 and the script-produced ADV-chrome flag G[0x6c1]=1 are present at that boundary. --scene SC0000 --boot remains available only as the explicit single-scene diagnostic harness. This lands the boot/title/New Game entry half of B1–B3; proving a completed scene return plus boundary cleanup still belongs to B1 completion.

TITLE SFX packed-raw correction (2026-07-20). Hover and activation callbacks were already executing their scripted 0xb5 starts. The load failed earlier because op 0xb4 uses universal packed SYS4INI/AAI ids, while Godot treated them as active-script manifest ids. The new packed-raw resolver maps TITLE's 0x2aea/SE020.WAV hover, 0x3321/SE015.WAV activation, and GAMESTART's 0x2aeb/SE013.WAV cancel through the existing channel players. A synchronized TITLE→GAMESTART→TITLE trace records every load/start with its filename, and manual validation confirms they are audible; BGM remains unaffected.

Pre-title video sequence implemented (2026-07-20). SYSTEM4 already owns the native sequence; the port did not lose an executable-side launcher. Its sole op 0x130 call returns an engine initial-root flag that is one at context construction and cleared only when op 0x9 resets/reloads root script id zero. SYSTEM4 calls LOGO.BIN and OP.BIN only while that flag is nonzero. The former stubbed-zero output explained the direct jump to TITLE. LOGO and OP then use the modal movie op 0x20f with raw catalog movies 0x335f/LOGO.AGF and 0x3364/OP.AGF; existing 0x236 is the distinct non-modal, scene-local movie-to-surface path. The VM now models the initial-root flag and clears it at the op-0x9 whole-stack root-reload boundary. Godot resolves a typed raw MPEG asset, reuses the asynchronous decoder and retained compositor, and parks the VM until EOF or mouse/Accept/Cancel input. Focused natural-boot tests prove LOGO -> OP -> INIT -> TITLE ordering and exact movie operands. MPEG audio remains explicitly deferred until the decoder abstraction has an engine-owned synchronized audio/volume contract.

Stage B1 — Persistent session and scene coordinator

Replace the single-SC0000-root assumption with an application-owned session that runs SYSTEM4 as its root. SYSTEM4's computed call-script is the authoritative scene coordinator: child scenes return to that frame, while globals, the host, and intentional retained state remain owned by the same live VM session.

Root-reload boundary implemented (2026-07-20). Ordinary op 0x2 child exits still return to their calling SYSTEM4 frame. Op 0x9 is the distinct native reset path: it discards the complete active script stack, clears scene-owned graphics/input/ADV state, cancels deferred SFX starts while preserving active audio, preserves global banks and process-owned host state, and starts raw script resource zero (SYSTEM4.BIN) at offset zero. The implementation propagates the boundary through nested calls without running any caller continuation and records the new root frame with FrameCause.RootReload. Native RE and the one intentional history-lifetime exception are documented in docs/engine-re.md; the history backlog remains preserved until its ownership is proven rather than guessed.

Required responsibilities:

• Own global integer/string banks and any proven external/profile state across scenes.
• Preserve the SYSTEM4 root while distinguishing ordinary child frames from scene-boundary children for diagnostics and lifecycle assertions; do not perform host-driven top-level replacement.
• Define scene-owned versus session-owned host state and tear each down at the correct boundary.
• Preserve intentional system-owned surfaces, configuration, and audio while releasing scene-local state.
• Expose deterministic transition evidence: outgoing scene, reason/decision, incoming scene, and state summary suitable for tests.

Completion evidence now present: SYSTEM4 reaches computed child scripts in one VM; ordinary children return to SYSTEM4; op 0x9 performs a tested whole-stack reload of SYSTEM4; selected globals and process-owned state survive; and scene-owned presentation/input state is released. Manual validation of a natural gameplay route through the first 0x9 remains deferred: Himegari's readily accessible return-to-title choice belongs to the still-unimplemented frontend exit-request policy, while the other known natural paths require later gameplay, game over, or completion. Do not use TITLE's post-0x1 developer menu as evidence; native 0x1 is non-returning, and the port now propagates that exit request instead of falling through into the hidden bytecode. See docs/engine-re.md.

Godot debug scene launcher (2026-07-20; implemented and manually validated). The first version is deliberately narrower than arbitrary hot swapping:

• Expose an F4-style Godot overlay only while TITLE.BIN is the persistent VM's active SYSTEM4 child.
• Resolve the chosen .BIN through the existing SYS4 catalog, then ask the VM to return the current TITLE child frame with the game-authored coordinator writes (G[0]=1, G[0xaba5c]=-1, G[0x62ccf]=0, and selected packed id in G[0x699]) applied on the VM thread.
• Let SYSTEM4 resume at 0x2b0 and execute its real entry wrapper and computed call-script; do not replace the VM root or call the selected scene directly from Godot.
• Disable switching while another scene is active. That scene must reach its own terminal cleanup and then either return through SYSTEM4's post-child cleanup or execute its genuine op 0x9. A separate clean relaunch remains the escape hatch for a stuck/incomplete scene.

The runtime now has a generic debug-only "return this exact active child frame with queued global writes" request, thread-safe frame-generation/stack reporting, and the distinct DebugReturned trace outcome. TITLE does not park in ADV op 0x72: its visible menu continuously polls input and executes a 1 ms op-0xc8 sleep at TITLE@0xe5. The request therefore targets the observed active frame generation and is consumed by the VM thread at its next completed opcode boundary, before another TITLE opcode can execute. SignalInput is used only if the target happens to be in a real ADV wait, avoiding a stale signal that could advance the selected child. Synthetic coordinator tests cover both an ADV wait and TITLE's sleep/poll shape, selected- child dispatch, ordinary SYSTEM4 continuation, stale/ineligible request rejection, and selected-child op-0x9 whole-stack propagation.

This launcher would provide the real visible TITLE→selected scene sequence and preserve the coordinator boundary, but it cannot manufacture valid late-game state. The current direct harness and opcode coverage suggest early ADV scenes and DEBUG.BIN are plausible targets; later scenarios, GAMECLEAR, battle/map, and profile-dependent scripts may still require progression data or missing opcodes. A startup-only/direct-scene selector is cheaper, but it is merely a UI for --scene ... --boot and provides no transition-lifecycle evidence. An unrestricted in-process switch would additionally require VM cancellation, task joining, movie/audio disposal, locator/trace regeneration, and an explicit global-state policy, so it is not a quick or trustworthy first version.

Menu population and selection contract. The runtime SYS4 catalog—not build/ inventory—is the source of truth. Himegari currently has 481 unique base-catalog .BIN records: 136 SC####, 164 SP*, 8 DEBUG*, 29 initializer-named scripts, and 144 other named scripts. Each menu row keeps the packed resource id as its identity and carries display name, pack selector, raw index, archive, size, and category; names are labels rather than keys so future append-pack collisions remain representable. Population should enumerate base Catalog.Files plus every mounted append catalog, exclude placeholders/non-BIN records, and compute packed_id = (pack_id << 24) | raw_index without parsing all scripts up front. The selected script is decoded and validated only when Launch is pressed; an unsupported decode reports an error and leaves TITLE running. The currently mounted append pack contributes 39 additional .BIN records, so the shipped launcher smoke test sees 520 distinct packed script ids.

The initial UI groups entries rather than implying every BIN is a standalone scene:

• Scenario: SC####.BIN, naturally sorted by number.
• Secondary/event: SP*.BIN, naturally sorted by name and suffix.
• Debug: DEBUG*.BIN.
• Other/expert: every remaining script; the separate All filter includes every category. Initializers, callbacks, data routines, and modal UI scripts may require caller-owned state and may immediately return or corrupt the live session.

SYSTEM4.BIN and TITLE.BIN are not launchable in the first version; recursively dispatching either through SYSTEM4's child slot is not a scene test. Search is case-insensitive over name and hexadecimal/decimal packed id. The detail pane shows name, category, packed/raw id, archive, size, and the fixed warning that launch uses the current live global/profile state. Compatibility or opcode-gap badges are deferred until coverage logic has an engine-owned runtime API; the menu must not parse generated Markdown or call Python tooling.

The implementation should leave one explicit extension point for future test sequences: DebugLaunchPreset(label, packed_script_id, extra_global_writes, note). Catalog rows use only the four coordinator writes above; profile-authored presets may later add proven story/progression globals without turning the menu into a free-form state editor or save backend. Arbitrary PC/offset jumps are out of scope.

Implementation order. (1) Add catalog script-entry enumeration with packed ids and unit coverage for base/append mounts, placeholders, duplicate names across packs, and category/sort/filter behavior. (2) Add a generic VM debug request targeted at an exact active frame generation; it applies an immutable set of global writes on the VM thread and returns that child at the next opcode boundary. Test SYSTEM4→TITLE→selected child, ordinary child return/cleanup continuation, op-0x9 propagation, TITLE's sleep/poll loop, and stale/ineligible request rejection. (3) Add the Godot F4 overlay (PopupPanel, search/category controls, ItemList, detail pane, Launch/Cancel), consume all overlay input, and enable Launch only for the active SYSTEM4 > TITLE stack. (4) Add a Godot smoke test for catalog population and request wiring, then manually validate TITLE -> DEBUG -> 0x9 -> SYSTEM4 -> TITLE before expanding the selectable categories or adding presets.

Steps 1–4 are complete. F4 opens the Godot PopupPanel only for the exact active SYSTEM4.BIN > TITLE.BIN stack; search, category filters, packed-id metadata, guarded Launch, and Cancel are live. SYSTEM4.BIN and TITLE.BIN remain visible but unlaunchable. While the panel is open, AGE gameplay input is not forwarded. Launch reparses the selected packed id before queuing any writes. The threaded Godot selftest constructs the catalog and panel and currently reports 520 unique packed scripts. Manual validation confirmed TITLE -> F4 -> DEBUG.BIN: its four scripted ADV pages at 0xc7, 0x110, 0x17b, and 0x1ed were presented, its terminal op 0x9 at 0x1fb ran, and SYSTEM4 reconstructed the visible TITLE menu. No launcher/session-lifecycle discrepancy was observed. DEBUG-specific content oddities are not acceptance failures for this developer route and remain out of scope unless they reproduce in a normal game script.

Stage B2 — Faithful full boot

Replace --boot's diagnostic seeding and separately injected inherited surfaces with normal boot execution. First prove the installed game's actual ordering; do not assume the current helper lists are complete.

The boot path must cover two existing categories:

• System/session initialization currently approximated by INITCONFIG, INIT2, and INIT, including host-visible side effects that CaptureHost discards.
• Game-data initialization represented by the *INIT family used by the headless boot/session tools.

Completion evidence:

• A fresh application reaches the same initial title state without --boot, --seed, or manual surface injection.
• Required globals come from executed scripts or clearly identified profile/native defaults.
• Inherited retained state has a traceable owner and lifecycle.
• A boot snapshot is reproducible for tests, but the shipped path performs the real boot rather than loading a developer snapshot.

This stage may reveal platform/install-selection work; track that separately in docs/platform-portability.md rather than folding cross-platform export into Phase B.

Stage B3 — Title and New Game happy path

Implement only enough menu behavior to choose New Game naturally and enter the story. This is primarily a state-initialization and dispatch slice, not a mandate to complete every submenu.

In scope:

• Title/main-menu presentation required by the executed path.
• Keyboard/mouse selection and the menu-specific coroutine/hotspot forms actually reached.
• Configuration defaults that affect New Game or the subsequent runtime.
• New Game initialization and its transition request.
• Natural empty-save-state behavior when no saves exist.

Deferred to bounded follow-ups unless the happy path requires them:

• Full configuration UI and every setting.
• Load/save implementation and save-format reversal.
• Extras, galleries, replay modes, and unrelated submenus.
• Menu visual polish that does not obstruct correct selection or state production.

Completion evidence: launching the application, selecting New Game, and reaching SC0000 with no manual state seeds. The resulting SC0000 opening state must match the Phase A visual/audio baseline.

Stage B4 — Natural progression through SC0000

Run the completed scene inside the persistent session and honor its real terminal transition. This stage closes the currently unidentified decision-to-scene boundary described in docs/scjump-progression.md.

Completion evidence:

• Title/New Game reaches SC0000 through actual script/native dispatch.
• SC0000 completes without the developer auto-advance harness.
• The outgoing decision, selected next script, and persistent global changes agree with the original run.
• The first dungeon scene starts without reconstructing the VM or reseeding state.

Stage B5 — First-dungeon vertical slice

Do not scope “gameplay” as maps + units + items + magic + combat + AI + win/loss all at once. After B0 identifies the first real interaction, select the smallest end-to-end loop that exercises authoritative script state. A likely target is:

• Load and render the first map and its initial UI.
• Populate the units required for the opening state.
• Select one unit and display its relevant state.
• Perform one legal move or scripted action.
• Resolve one combat or event interaction if the natural opening reaches one.
• End one player action/turn and return to a stable input boundary.

The exact acceptance path must follow the installed game's first dungeon rather than forcing this example shape. Items, skills, magic, AI, win/loss, deployment, and progression are added only as the selected path requires them.

Completion evidence combines original-game observation, executed-opcode/call traces, visible map/UI output, and before/after global-state comparisons for the action.

Later Phase B breadth

Once the natural spine and first gameplay loop are trustworthy, broaden in independent tracks:

• Remaining title/configuration/load/save branches.
• Save-file format and restoration of a persistent session.
• Map navigation, camera, terrain, deployment, and turn lifecycle.
• Unit statistics, equipment, inventory, skills, magic, heroine forms, and progression.
• Combat resolution presentation, enemy turns/AI services, and win/loss transitions.
• A full chapter of ADV and the scene types encountered between gameplay segments.
• STINIT and other data schemas when their runtime consumers make them necessary.

The high-level Phase B direction remains canonical in docs/remake-architecture-and-roadmap.md; this file only provides the execution framework.

Slice record template

When a stage becomes active, add its concrete plan/results to this document or the then-current slice plan without pre-planning all later systems. Record:

• Starting state and exact reproduction path.
• One observable end condition.
• Executed unknown/effectful opcode families.
• Required global/profile/host state and its producer.
• Native/original evidence to capture.
• Explicit non-goals.
• Automated regression gates and manual acceptance check.
• Result, remaining discrepancies, and the next decision gate.

Decision gates

• After B0: confirm or revise the boot/title/SC0000/dungeon sequence using observed load order.
• After B2: decide whether session state is trustworthy enough to remove the diagnostic boot path from ordinary runs.
• After B3: choose whether save/config work blocks natural SC0000 entry; otherwise defer it.
• After B4: use the actual first-dungeon trace to write the concrete B5 slice rather than guessing its systems in advance.
• After B5: choose breadth based on the next natural blocker, not opcode coverage percentage alone.