new launcher and save converts and pipeline

This commit is contained in:
bryanthaboi
2026-07-25 08:34:49 -04:00
parent 6625391f76
commit 9307488dc1
58 changed files with 10971 additions and 348 deletions
+149
View File
@@ -0,0 +1,149 @@
# Launcher
The launcher is `src/import/RomImporter.lua`, the first-run / title screen
that runs before `Game:load`. Besides ROM import (see the file's own header)
it hosts a tabbed shell covering per-game save slots and a mod manager. This
file documents the runtime model; the visual spec lives separately.
## Tab structure
`self.tab` is one of `"red"`, `"blue"`, `"yellow"`, `"mods"`. The tab bar
draws one chip per game plus a MODS chip and rebuilds `self.tabRects` every
frame so `mousepressed` can dispatch clicks; switching tabs mid-import is
allowed (a dropped ROM still routes by SHA-1 regardless of which tab shows).
- A game tab (`_drawGamePanel`) shows the ROM card, the SAVE FILES card, the
Play button, and the SAVE SLOT card in a responsive two-column grid (see
Responsiveness). The MODS tab (`_drawModsPanel`) shows the mod list instead.
- The self-updater banner (`self.Check`, see `docs/updater.md`) draws as a
centered pill in a reserved band just above the footer, on every tab. That
position is unchanged by this redesign, so `docs/updater.md` needed no edits.
## Save slot model
All slot I/O lives in `src/core/SaveData.lua` and goes through the same fs
abstraction (`persistFs`) every other save/options call uses, so portable
mode (an `io.*` filesystem used when `portable.txt` marks the install)
keeps working unchanged.
- **Files.** A version's playthroughs live under `saves/<version>/`, one file
per slot: `saves/<version>/slot1.lua` plus a rolling `.bak` and staged
`.tmp` witness (`slotNames`), mirroring the write/recovery discipline
`SaveData.save`/`load` already use for the flat legacy file. Slot ids match
`slot%d+`; `createSlot` allocates one past the highest existing number so a
reused id can never collide with a lingering file.
- **Registry.** The ordered slot list and which one is active persist in
`options.lua` (via the existing `SaveData.loadOptions`/`saveOptions`):
`options.saveSlots = { [version] = { list = {"slot1", ...}, active = "slot1" } }`.
- **Active slot resolution.** `saveNames(version)`, the function every
existing caller (`TitleState` hasSave/load/save, recovery order) already
goes through, now resolves the *active* slot instead of a fixed flat name.
Resolved once per version per process (`ensureVersionSlots`, cached in
`activeSlotCache`/`slotsChecked`): a registry entry wins; otherwise a lazy
legacy migration may create one; otherwise the flat legacy path is used
(`save.lua` / `save_blue.lua`), so a pre-slots install keeps working as before.
- **Legacy migration.** One-time per version, lazy on first
`listSlots`/`load`/`saveNames` call (`tryMigrateLegacy`): if a flat legacy
file exists and no `saves/<version>/` registry does, its main + `.bak` are
copied into `saves/<version>/slot1.lua(.bak)`, verified readable
(`decodeSlot`: main, then `.tmp`, then `.bak`), and only then are the
originals removed and `slot1` registered as active. A copy that fails to
verify leaves the originals in place; migration never loses data.
The launcher-facing API:
- `SaveData.listSlots(version)` -> array of `{id, exists, name, meta}` for
every registered slot. `name` is the save's player name, or `nil` for an
empty slot; `meta` is `{badges, timeText, dexCount}` (the same fields the
title screen's `ContinueInfo` shows) or `nil`. The pure part,
`SaveData.slotSummary(save)`, is unit-testable with no filesystem.
- `SaveData.setActiveSlot(version, slotId)` registers the id if new, persists
it as active, and updates the process cache so the very next save/load
lands there. The launcher calls this the moment a slot row is clicked
(`RomImporter:_selectSlot`); pressing Play needs no signature change, since
`Game.lua`/`main.lua` still just call `SaveData.load()`/`save()`.
- `SaveData.createSlot(version)` -> new slot id, registered but with **no
save file written**. An empty slot means the title screen offers NEW GAME
only, which needs no further changes.
## Launcher mod manager
`src/mods/LauncherMods.lua` is a launcher-only read of the mod set. It runs
before `Game:load`, so **it never loads a mod's entry chunk**; only
`manifest.json` is read and validated (`src/mods/Manifest.validate`), the way
`Loader:_discover` finds mods without running them. The real loader
(`src/mods/Loader.lua`) still owns the actual load at boot.
- `LauncherMods.list()` scans `mods/` one level deep (first id wins on a
duplicate) and returns one row per mod:
`{id, name, version, badge, description, enabled, status, statusDetail}`.
`badge` is the manifest's `category`, falling back to `profile`, then
`"MOD"`, uppercased. `enabled` reads `options.mods[id]` (missing means
enabled, matching the loader's own default).
- `status` is `"ok"`, `"warn"`, or `"conflict"`, computed by the pure
`LauncherMods.deriveList`/`statusFor` against `ManagerState.resolveToggle`
and the validated manifests: `conflict` when enabling this mod collides
with another enabled one; `warn` for an out-of-range `game_version` or an
absent/disabled/wrong-version hard dependency; `ok` otherwise. Having no
`love.*` calls, this half is table-driven by the test suite on its own.
- `LauncherMods.setEnabled(id, bool)` persists `options.mods[id]` as a plain
boolean, the exact shape `Loader:_saveState` writes, so the running game
and the in-game `ManagerState` see the change on next boot. The mods panel
calls this on every toggle and re-derives the list right away
(`RomImporter:_refreshMods`) so a status change (e.g. a new conflict)
shows without waiting for a reload.
- `LauncherMods.installZip(path)` mounts the archive with
`love.filesystem.mount`, locates the mod root via `locateRoot` (manifest at
the zip root, or inside one top-level folder), validates its manifest, and
copies the tree into the save-dir `mods/<id>/` before unmounting. Rejects a
duplicate of an already-installed mod id, and accepts either an external
path string or a LOVE `DroppedFile`, staging a dropped file into a save-dir
temp first (mount only reaches save-dir-relative paths), the same way
`RomImporter` handles a dropped ROM. A failed copy rolls its partial tree
back, and every path unmounts and clears the staged temp file.
## Import / Export save
The SAVE FILES card wires a raw Gen1 `.sav` battery image to the save slots
through `src/import/SaveFileIO.lua`, which sits on top of
`src/save_convert/SaveConvert.lua` and the slot API in `SaveData`.
- **Import save** is live once the game's ROM is imported (playable). It opens
a native `.sav` picker (`chooseSav`, the per-OS dialogs mirror `chooseZip`;
Android has no picker and shows a drop hint). `SaveFileIO.importToSlot`
reads the bytes (an absolute path, a dropped LOVE file, or raw bytes),
guards the 32768-byte size, runs `SaveConvert.importSav` (which also rejects
a bad main-data checksum), then registers a fresh slot (`SaveData.createSlot`),
writes it (`SaveData.writeSlot`), and makes it active (`SaveData.setActiveSlot`).
The meta stamp is re-stamped off `gen1_import` to the current numeric format
so `SaveData.load`'s migration pass accepts the slot. On success the SAVE SLOT
panel is refreshed with the new slot selected.
- **Export save** is live only when the active slot actually holds a save
(checked against `listSlots`). `SaveFileIO.exportActiveSlot` loads the active
slot, encodes it back with `SaveConvert.exportSav` (a slot never keeps
`rawImport`, so this is a zero-filled template export, which is valid), and
writes `exports/gen1recomp-<version>-<slotId>.sav` in the save directory
(`love.filesystem.createDirectory("exports")`). It returns the absolute path
(`love.filesystem.getSaveDirectory()`), which the notice line shows with a
desktop "Open folder" affordance (`love.system.openURL("file://" .. dir)`).
- **Drag-drop.** `filedropped` routes a `.sav` to the import path for the
currently active game tab; when a non-game tab (mods, or the locked yellow
placeholder) is showing it defaults to red, the always-present first game
(`_savedropTarget`). `.gb` (ROM) and `.zip` (mod) routing is unchanged.
- **Failure UX.** Every error path (wrong size, bad checksum, write failure,
nothing to export, ROM not imported yet) surfaces as a red notice line on the
card. Nothing raises and nothing silently no-ops.
`SaveFileIO` is love-free enough to unit-test through the same in-memory
filesystem stub the slot backend uses (`tests/engine/save_file_io_tests.lua`).
## Responsiveness
Every measurement derives from `love.graphics.getDimensions()` each frame
plus the existing global scale `s = clamp(height / 768, 0.7, 1.6)`; nothing
assumes a fixed window size. The game panel's two-column grid (ROM/SAVE
FILES/Play on the left, SAVE SLOT on the right) collapses to one stacked
column, slot card below Play, when the window is too narrow for both
`~300 * s`-wide columns. The save-slot list and the mod list both scroll
(wheel, or drag on touch/desktop) clamped to their own content extent,
recomputed every draw. The tab bar labels only the active chip so it stays
narrow-safe, and content caps out at `~1440 * s` wide, centered.
+105
View File
@@ -18,6 +18,111 @@ Regenerate the reference straight into a wiki checkout:
luajit tools/gen_registry_docs.lua ../pokemon-gen1-recomp-project.wiki
```
## Rendering pipelines
Most registries hand the engine *content*. `render_pipelines` hands it
*drawing*: a pipeline is a display mode a mod owns, which may replace the
overworld's world pass with geometry of its own and/or post-process the
finished image. `mods/voxel_world` is the worked example — a 3D diorama
overworld plus a tilt-shift miniature pass, in about 120 lines of glue over
its renderer.
A record declares what the mode *is*; the engine
(`src/render/Pipelines.lua`) supplies everything about *being a display
mode*: the OFF/1/2/3 ladder, an options row next to TILT, a hotkey,
persistence in `save.options.pipelines`, and the rule that a world pipeline
and the engine's own TILT are mutually exclusive.
```lua
mod.content.render_pipelines:register("diorama", {
label = "DIORAMA", -- options row label
levels = { "OFF", "15", "35", "50" }, -- ladder; defaults to OFF/ON
hotkey = "6", -- checked after the engine's keys
priority = 20, -- highest eligible wins the world
available = function() return Renderer3D.ok() end,
update = function(dt, level) Camera.ease(dt, level) end,
drawWorld = function(ctx) return renderScene(ctx) end,
})
```
Three draw stages, each optional; a record needs at least one:
| stage | signature | runs |
| --- | --- | --- |
| `drawWorld` | `(ctx) -> canvas \| nil` | instead of the flat/tilt world pass |
| `worldPresent` | `(canvas, ctx) -> canvas` | over the world, **before** the UI composites |
| `present` | `(canvas, ctx) -> canvas` | over the whole frame, world and UI alike |
`worldPresent` is the one to reach for when an effect must leave dialog
boxes and menus crisp — a depth-of-field or colour grade on the world only.
`present` is for effects that genuinely own the screen, like a CRT curve.
`ctx` carries the frame: `state`, `cam`, `vw`/`vh` (world-pixel view),
`width`/`height` (window pixels), `scale`, `level`, `paletteFor(map)` and
`spriteColors(map)`. It also carries `ctx.drawFx(project, scale)` — call it
with your own projection and the engine draws every active field effect
(the "!" bubble, the Poké Center heal machine, the Fly bird, the fishing
rod, Rock Tunnel darkness) at its correct anchor under your camera. There
is exactly one copy of each effect, so a new engine effect works in your
pipeline without you touching anything.
Three rules worth knowing:
- **`gate` governs input, never the draw.** It decides whether the player
may *change* the mode (default: free-roam overworld only). A mode that
stopped rendering during a warp would flash the flat 2D world every time
the player walked through a door.
- **`available` is re-read every frame** and is the only thing that decides
whether the mode can render at all. Answer `false` on a headless run or a
driver with no depth canvas and the engine silently keeps the vanilla 2D
path — which is why shipping a pipeline enabled is safe.
- **A callback that throws retires its pipeline**, attributed to your mod in
the manager's error feed, and the frame falls back to 2D. A broken
renderer costs the player a display mode, never the game.
Returning `nil` from `drawWorld` is a normal answer meaning "not this
frame"; the engine draws the vanilla world instead.
## Battle sprite scaling
The enemy's front pic draws at 1x and the player's back pic at 2x, the way
the Game Boy did. A mod can override either, per species or per image.
Per species, on the `pokemon` record:
```lua
-- MEW's back pic renders 1.5x; its front pic is untouched
mod.content.pokemon:patch("MEW", { battleScaleBack = 1.5 })
```
`battleScaleFront` scales the enemy pic, `battleScaleBack` the player pic;
both take a number in `0.25 .. 4.0`.
Per image, on the `battle_sprite_scales` registry, keyed by the asset path
exactly as the data references it:
```lua
mod.content.battle_sprite_scales:register("abra_back", {
path = "assets/generated/battle/back/abrab.png",
scale = 1.5,
})
```
An image-level entry beats the species scale for that one pic, and it is
the only way to scale a pic that is not species-keyed — the player's
trainer back sprite, held on screen until "Go!", is a bare image path.
The resolution order at draw time is **image-level → species-level →
default** (1x front, 2x back).
- **The pic stays grounded at every scale.** The player pic keeps its feet
flush on the text-box top (`y = 96`); the enemy pic keeps its bottom edge
and horizontal centre pinned in its 7×7 slot. A larger pic grows upward
and outward from that anchor, never off the shelf.
- **Scaling composes with the send-out grow.** The `AnimateSendingOutMon`
ball-to-pic grow multiplies your scale through each stage, so a rescaled
mon still grows into place from the ball, grounded the whole way.
## Developer console
Boot with developer mode on to unlock the in-game console and hot-reload
+126
View File
@@ -0,0 +1,126 @@
# Updater
A fused build (`love.filesystem.isFused()` true) ships a bundled `game.love`
baked into the executable, but that bundled copy is only ever the *fallback*.
On every launch, before anything else runs, `Boot.run` (`src/update/Boot.lua`)
looks in the save directory's `updates/` folder for a downloaded
`gen1recomp-X.Y.Z.love` payload that is both strictly newer than the bundled
engine version and runnable on this shell. If one qualifies, it is mounted
over `/` (so its files win over the fused source for every subsequent
`require`) and chainloaded in place: the payload's `main.lua` and `love.load`
run as if they had shipped in the executable. A dev/source checkout is never
fused, so `Boot.run` no-ops there and the working tree always runs itself.
The pieces are deliberately layered so the risky part is small. `Boot.select`
is a pure function (no `love.*` calls) that, given probed candidates and the
bundled `engine`/`shell`, decides what to run and what stale payloads to
delete. `Boot.probePayload` mounts one archive at an isolated mountpoint and
reads its `src/core/Version.lua` with `loadstring` (never `require`, so it is
never cached as a module) to learn its `engine` and `minShell`. `Boot.run`
orchestrates the crash guard, enumeration, selection, and the mount +
chainload, with full rollback on any failure. Checking for and fetching a
new payload is a separate, slower path: `src/update/Check.lua` is a thin
main-thread state machine the launcher screen polls, while the curl calls,
JSON parsing, and sha256 verification run on a background `love.thread`
(`src/update/check_worker.lua`) so a hung network call never blocks a frame.
## Version.lua fields
`src/core/Version.lua` carries three fields the updater reads directly (the
existing `modApi`, `linkProtocol`, `saveFormat`, and `cache` fields are
untouched):
- `engine` - the semver release, e.g. `"1.4.0"`. The repo default is the
`"0.0.0-dev"` placeholder; CI stamps the real `X.Y.Z` into the packed
`game.love` only, never the working tree. A `"0.0.0-dev"` engine always
reports itself up to date (it never chases a release, and it never counts
as a valid payload to chainload).
- `shell` - the native-shell contract this build's fused executable
implements.
- `minShell` - the lowest shell contract required to *run* this payload.
Bump `minShell` only when a payload needs something the currently-shipped
native shell cannot provide, for example a LOVE version bump, a new required
system binary, or a change to `love.run` itself (see Known limitations
below). An older shell refuses to chainload a payload whose `minShell`
exceeds the shell it provides; `Boot.select` keeps that payload in `updates/`
rather than deleting it, in case a future shell upgrade can run it, and
`Check`'s worker reports `needs_full` so the player is pointed at a full
installer instead. Do not bump `minShell` for an ordinary Lua/data release;
that is exactly the case the updater exists to avoid a reinstall for.
## Release assets
Each tagged release `vX.Y.Z` carries the existing per-platform archives
(`gen1recomp-X.Y.Z-macos.zip`, `-windows.zip`, `-linux.zip`,
`-android.apk`) plus two assets the updater itself consumes:
- `gen1recomp-X.Y.Z.love` - the payload, matched by the exact pattern
`gen1recomp-<version>.love` (see `isPayloadName` in `Boot.lua` and
`Check.parseRelease`).
- `sha256sums.txt` - `shasum -a 256` output (`<hex> <filename>`, bare
filenames) covering at least the `.love` payload. `Check.parseSums`
tolerates a leading `*` binary marker and a `./` prefix but expects the
filename otherwise to match the asset name exactly.
A release missing either asset is treated as "no in-place update available":
`Check` reports `needs_full` and sends the player to `Check.releaseUrl()`
(`https://github.com/bryanthaboi/pokemon-gen1-recomp-project/releases/latest`).
## Save-directory layout
Under the save directory (identity `pokemon-love2d`):
```
updates/gen1recomp-<X.Y.Z>.love downloaded payload(s)
updates/pending.txt crash-guard marker
```
`pending.txt` holds the filename of the payload currently being chainloaded.
`Boot.run`'s `chainload` writes it immediately before mounting, and removes it
on both a successful handoff and a clean rollback. If it is still present the
*next* time `Boot.run` starts, the previous boot died mid-handoff, so that
named payload is distrusted: it and the marker are deleted before candidates
are enumerated. Boot may still fall back to an older valid payload, or to the
bundled game, in that case.
## Update flow
1. **Boot** (every launch, fused builds only): crash-guard check, enumerate
and probe every `updates/*.love`, pick the highest engine that is
strictly newer than the bundled one and whose `minShell` this shell
satisfies, delete stale payloads, chainload the winner (or run the
bundled game if none qualifies).
2. **Check** (launcher screen): `Check.start()` kicks off an async check
against the GitHub releases API; safe to call every frame, it is a no-op
once a check is in flight or has reached a terminal state. `Check.state()`
reports `idle | checking | uptodate | available | downloading | ready |
needs_full | error` plus the latest version and download progress.
3. **Download + verify**: on `available`, `Check.download()` tells the
worker to fetch the payload, polling the growing `.part` file for
progress. On completion the worker re-fetches `sha256sums.txt`, verifies
the payload's sha256, and probes it with `Boot.probePayload` to gate its
`minShell` against this shell's `shell`. A verified, runnable payload is
renamed into place and reported as `ready`; anything else reports
`error` or `needs_full` and leaves `updates/` clean.
4. **Restart to apply**: a `ready` payload just sits in `updates/` until the
player relaunches; the next launch's Boot step (1) is what actually
mounts and runs it. There is no in-session hot-swap.
## Known limitations
- **`love.run` persists across handoff.** By the time `chainload` runs, the
bundled `love.run` has already returned its stepper to LOVE; redefining the
global `love.run` from the payload's `main.lua` does not affect the loop
already driving the frame. A payload that must change `love.run` itself
needs a `minShell` bump so an older shell refuses to chainload it rather
than running with half its intended behavior.
- **Android has no in-app download transport yet.** `check_worker.lua`
shells out to curl for both the release check and the download; curl is
absent on Android, so `Check` degrades to `status = "error"` there (the
launcher UI hides on that status) and the player is directed to the
releases page via `Check.releaseUrl()` instead.
- **Dev/source runs never self-update.** `Boot.run` returns immediately when
`love.filesystem.isFused()` is false, and a working tree's `engine` is the
`"0.0.0-dev"` placeholder that always reports up to date, so a source
checkout is always "the game" itself; updating it means pulling the repo.