Threads open a fresh Lua state with a full standard library, so they
stayed blocked wholesale. PotatoVoxel's prebuilder wants to run its
pure geometry phase on worker threads; the mod declares the new
"compute" permission and the sandbox hands out love.thread only
then. The worker runs the mod's own source (source-only, like every
mod file) and receives data only through channels.
Android ships no curl and the JNI bridge was GET-only, so mod.postLog
failed there with 'no POST transport on this platform' (HostShell.lua).
Add the mirror of httpDownload: GameActivity.httpPost (https-only,
hand-followed redirects re-POSTing the body, one-way), the JNI bridge
with the same old-APK-skew tolerance, the love.system.httpPost binding,
and the HostShell arm that rides it when curl is absent. The body
crosses the JNI as raw bytes (jbyteArray) so a log ring with arbitrary
UTF-8 cannot corrupt through modified-UTF-8 jstring conversion.
Font.drawBox leaves the caller's color at white; every other screen
resets to black right after calling it, but ManagerState.lua's draw()
and drawOverlay() never did. That's invisible on the vanilla tile font
(tile glyphs are black-on-transparent regardless of color) but renders
fully invisible white-on-white text once a mod's TTF font is active.
The targeted test only covered the RBY side (ItemEffects.lua,
TrainerAI.lua); src/battle/gen2/Battle.lua's EFFECT_LIGHT_SCREEN/
EFFECT_REFLECT fix had no test at all, spotted when asked whether the
Gold changes were covered.
Adds a minimal MACHOP/TACKLE gen2 fixture (same shape as
tests/gen2_move_effects_test.lua's), calls both MOVE_EFFECTS directly,
and checks a loaded catalog reaches the whole message template (Gold
wraps the full sentence, not just the stat name). Confirmed it catches
the regression: reverting Battle.lua to its pre-fix state fails 2 of
the suite's now 8 checks.
No existing test could tell a translated stat name apart from a raw
stat:upper() that never went through Strings() at all: every rose!
message assertion in the suite runs with no catalog loaded, where
Strings() is an identity function either way.
Loads a real catalog (Strings.load) that translates one stat name at a
time and checks it actually reaches the X-item, vitamin, and AI-trainer
X-item messages -- ROM-free, over tests/fixture_data. Confirmed this
catches the regression: reverting src/inventory/ItemEffects.lua and
src/battle/TrainerAI.lua to their pre-fix state fails 5 of 6 checks.
Strings(stat:upper()) is a dynamic argument -- tools/modkit.py's
STRINGS_CALL harvester only matches a literal string right after
Strings(/Strings.source(, so it can't discover "ATTACK"/"DEFENSE"/etc.
from these call sites. Translation coverage happened to still work
only because the same literals are independently harvested from
unrelated call sites (MoveEffects.lua's STAT_LABEL, BattleState.lua's
literal Strings("ATTACK") calls) -- real but fragile, found in review.
Reuse the codebase's existing pattern for exactly this situation
(MoveEffects.lua's STAT_LABEL): a local table built at require time
with Strings.source(...), which the harvester can see, resolved to a
translated label at use time with Strings(TABLE[key]). Adds one such
table to ItemEffects.lua (covering its X-item and vitamin call sites,
including "hp") and one to TrainerAI.lua.
Same theme as the RBY fix, found while checking whether Gold had the
same gap: EFFECT_LIGHT_SCREEN and EFFECT_REFLECT built their "'s
SPCL.DEF/DEFENSE rose!" message by raw string concatenation, bypassing
Strings() entirely -- unlike most other messages in this file (e.g.
"%s\nused %s!" a few lines up), which already go through it.
Wrap the whole message template in Strings(), matching that existing
pattern; the substituted name still comes from monName() as before.
Gold's gen2/Battle.lua has many more messages built the same
unwrapped way (fainted!, learned..., missed!, and so on) -- that is
the much larger "Battle messages" gap already tracked separately and
deliberately left out of this change.