GitHub #388 (claim 1). Every S1 alchemy unit applied its DEF buff under one shared key ('thorn2:bastion'), so a second
unit over the same ally replaced the first one's buff (two units: DEF +60, not +120; a normal and an elite: +75, not
+135) while both units' heals landed. PRTS 引星棘刺 技能1 备注: 「效果均可叠加」. Each alchemy unit now has its own
DEF buff key (assigned at its first tick), so the buffs add (aggregateMods sums defFlat) and each lapses with its own
unit. S3's 不叠加 debuff is untouched.
Golden: roster-033 (a normal 引星棘刺 with S1, pass 1) and roster-034 (an elite one, pass 1) move — one operator's
consecutive S1 units overlap for a few seconds (6 / 7 s + 3 s 心相 against a 7 / 6 SP cost) and now add their DEF on
an ally both cover (033: killed 46 → 64, leaks 121 → 103; 034: damage taken by 5_13_b 41498 → 14903 …).
Tests: test/content/kits_alt_t5.test.js (+1: normal + elite on one ally = +135, the longer unit's +75 stays alone,
nothing afterwards; fails on the parent with +75).
GitHub PR #281 by @IceCodeNew (this commit is substantially the PR, ported onto 0.2.2-dev; the field-detail range
idea after @BBleae's fork commits c6437f1 / b451bad, credited here instead of the PR's Co-authored-by trailers). The
unofficial skill-range preview the PR first had is gone (the owner's reply of 2026-10-07: only official-like
behaviour); what remains is the range the unit fights with now.
screens/game/range.js inspectRange resolves the open card's unit: an own board piece in prep (its loadout or
stand-in record, its facing — this replaces the prep selection's range effect); in battle or on a scouted board an
ally op / token on the field, at the snapshot's tile, with the live stats' range when there is one; never an enemy,
a dead or undeployed unit, or a scouted bench unit (UnitInfo `area` hand / temp — the tag the server sends since the
#385 fix; renderInfo now keeps it). shared/protocol.js unitStatsEntry gains an optional `dir` (the unit's facing now:
the snapshot tuples carry none and UnitInfo.dir is the facing at send time). tokenVariantFor moves to
gameLogic/loadout.js. Drag, the wheel and the pen clear the overlay.
Added on top of the PR: a test from the server's own scout m.field (board operator ranged, hand / temp none), docs
(PLAYING.md, design/client.md, design/network.md), comments.
Tests: test/ui/inspect-range.test.js (6), test/ui/playtest6_summons.test.js, test/match/unitStats.test.js;
test/ui/inspect-range.e2e.test.js (SP_E2E + CHROME_PATH: run here, 1/1). Golden unchanged (display; m.unitStats only).
GitHub #394 / PR #395 by @Xinglan233. The wheel's capture listener committed the previewed direction on Enter whatever
had the focus, so Tab to 准备 / 设置 / the exit button / ✕ and Enter deployed (the PR let only a focused ✕ keep its
native Enter). Now gameLogic facingEnter decides: Enter on ✕ cancels; on any other control (button, link, form field,
ARIA control) it is swallowed — the wheel already holds Space and the shortcuts; elsewhere it commits the previewed
direction, and does nothing while none is previewed. The wheel (role="dialog") takes the focus when it opens
(tabIndex -1, no focus ring over the screen), so a button focused before the drop cannot swallow the confirming Enter
and the first Tab reaches ✕; nothing restores the old focus when it closes (a restored 准备 would take the next key).
Docs: PLAYING.md's key paragraph. Tests: test/ui/feedback5-hotkeys.test.js (facingEnter and its wiring, 2 new);
test/ui/facing.e2e.test.js (SP_E2E, the PR's drag / Tab flow, 4 cases: run here in headless Chrome, 4/4 pass, 4/4 fail
on the parent — Enter on the focused exit button deployed). Golden unchanged (client only).
GitHub #385 / PR #387 by @zasx77771. The prep scout sends the hand and the 临时整备区 as units, and ownerBoard put every
op unit on the popup's synthetic board, so the popup marked bench operators 在场 and its 成员 header disagreed with the
server's ring (炎: ring 1, header 3; 远见: ring 2, header 3 with the temp copy).
The scout's units now carry `area` 'board' | 'hand' | 'temp' (the server tags from PR #387; the tag survives the
boss-half remap, the row does not). ownerBoard sorts them into board / hand / temp instead of dropping the bench as
the PR did: dropping the hand would break the bonds that count the 整备区 (投资人 / 远见 / 奇迹 — a hand-only 远见
member: ring 1, header 0). bondMembers / memberHeadCount then count exactly as the player's own popup: hand = owned
and in hand, temp = owned only. A unit without an area (battles, older servers) stays on the board.
Docs: PLAYING.md's 观看谁 paragraph (the bench is listed in prep); comments in bondStrip.js / game.js.
Tests: test/ui/watch-bonds.test.js (36; 2 fail on the parent), test/match/prep-bench.test.js +
feedback5-boss-pair-prep.test.js (area tags; 4 fail on the parent). Golden unchanged (views only).
GitHub #389 / PR #390 by @Xinglan233 (this commit is substantially the PR: the snapshot and its test file, ported).
onArt read target.items after grantChess; a gain that completes a merge runs _mergeChess first, which consumes the
target and empties its items, so the copy loop ran on an empty list (two base 维式重锤 never became the +25% one).
The item ids are now recorded before the gain.
Added on top of the PR: after a promotion the copied items are never equipped onto the elite (an already-advanced
item used to go straight back onto it via equipDirect). The promotion unequips the copies (PRTS 帮助 自动卸除 on
合并) and the 画卷 note says the obtained equipment arrives unequipped, so they wait in the hand; a copied normal item
merges with the returned original as before. The non-promotion path is unchanged.
Tests: test/match/scroll-copy-merge.test.js (8; 4 fail on the parent, the advanced-item case also fails with the
PR's snapshot alone). Golden unchanged (no scenario casts 画卷 into a promotion).
- Conflicts: the stats entry (merged fb7-stats) and the settings button both sit in the lobby / room top bars — both kept;
public/js/ui/settings.js (#321 diagnostics, voiceLang, SettingsButton); docs/PLAYING.md table rows resolved line by
line against the merge base (#368, #354, #333 edits kept); docs/SIM.md / design/network.md snapshot field lists
(resolved + ammo / wolves / neg / stand / standCut); design/client.md Screens line (#368's AI switch kept).
- Grok review: MERGE AFTER FIXES — the shared-field capsule finding fixed on the branch (1a880699).
- Golden: roster 23, bonds 2, fields 11, matches 18, diy 11 — #308 totals / snapshots and #333's shop RNG (union OK);
broad node --test 5473 (5472 pass, 0 fail, 1 skipped); lint 0 errors; typecheck; i18n strict clean.
A push and a kill on the same sim step bring the displace fx and the 'die' together: the view died before its slide
could anchor, and render/app.js syncBattle gives a dying view only its position (v.x = s.x) — never sync(), the only
place that set slide.live — so update() held the slide's start until its 0.5 s timeout, past the 0.8 s DIE window,
and dropped it: the body stayed on its pre-push tile (8.852 instead of the sim's 10.5; a summoned enemy and a flyer
alike). The per-view step of syncBattle is now units.js syncView (unchanged for living, device and revived views); its
dying branch calls UnitView.followSample, the position part of sync(): the sample anchors a slide from `at` on, and is
the position only when no slide owns it. A slide whose destination snapshot never shows (the unit left the snapshots)
now goes on to the fx's destination after the 0.5 s hold instead of being dropped where it started.
The frame-order test fed the view with sync() after its death and stepped the battle through the harness (whose
step() drained the events away from the test's own queue): it now steps the battle itself and goes through syncView.
Tests: test/render/unitview.test.js 38/38 — the app.js frame order over real battles: a 1.7-tile push; a kill on the
push's step for a route enemy, a summoned one and a flyer (the three failed before: bodies at 8.852 / 7.262 / 8.852)
and a kill 6 ticks later — the body ends where the sim's lies, every frame step < 0.35 tiles, no step back for the
same-step kills; a never-shown slide goes to the destination. Render interp, attack recovery, champagne, eyemask,
enemy attack clips, sim displace-fx: 94/94 with unitview.
Counting rule (ui/stats.js countsTowardStats, the owner's decision of 2026-10-08: do not count the matches entered and left
right away, but keep them in the history): a settled match (m.result arrived) always counts; a match given up (放弃模拟)
counts once at least one of the player's battles had finished — a round cleared (the server's own rule, platform.js
_quit: leaving while round k runs has passed k - 1, in its settlement k) or already eliminated. Anything earlier is
history only: marked 中途退出 and 不计入统计, dimmed, listed in 对局记录, left out of every aggregate (games, wins, win
rate, strategy table, titles, rounds, time, combat totals); 总览 says how many were left out.
- A leaver gets no m.result, so the page records the match itself: quitMatch (放弃模拟) → recordQuit builds an
`end: 'quit'` record from the app state and a tracker (observeMatch, installed in main.js) that remembers when the page
first saw the match and the last round the player was seen alive in. No record for a spectator seat, a match with its
settlement already there, a match the page never saw; matches the server ended and closed tabs are not recorded.
Defects of PR #323 as submitted, reproduced (scratchpad/w7-stats-notes/repro-pr-bugs.mjs, loads-check*.mjs) and fixed:
- the title screen loaded all 16 game data files (5.3 MB) at boot: <StatsHost/> called useGameData() at mount; now the
host renders nothing until opened, the overview reads config / bands / assets and the settlement re-view loads the rest
when opened (the title screen loads the same 3 files as before);
- importing a file put the list oldest first (the page shows the first 20 as the newest and the cap would drop the
newest); the list is newest first everywhere, also when loaded from storage;
- recordResult wrote over the envelope of a NEWER build (a downgrade, against the PR's own rule); it refuses now, as
saveStats does;
- 1000 four-player records are ~7.2 M characters (the PR's ~1.5 kB/record is a solo match) against a ~2.5 M-character
quota: MAX_RECORDS 1000 stays, plus a 800 000-character budget (the oldest go first), and only the newest 30 records
keep their rows (lineup, bonds, teammates' names; the replay needs them) — older ones are compacted to the player's own
summary row (~620 characters), a refused write retries with half the history; unknown fields and oversized values are
dropped (a strict schema), import files are bounded (8 MB);
- the re-view clobbered the app store (match.public / match.result, the route, the music); ResultScreen is split into
ResultScreen (the store) and ResultView (props: result, myId, back label), so the stats page renders a stored match
itself; a 自选 pick and a 补位 stand-in were dropped from the stored lineup (the replayed card drew the plain
operator); DIFFICULTY_NAMES were not translated; a band cell was a display:flex <td>; the overlay stayed open when a
co-op match started; the 清空 two-step blink was replaced by the shared confirm dialog.
- Corrupt storage: unreadable data is moved to sp.pref.stats.damaged and the page says so; damaged rows are dropped and
counted; duplicates, an oversized or unsorted list are repaired on load; a pre-`end` (PR-era) envelope reads as settled.
UI: a full-screen overlay in the grammar of 干员调配 (top bar with 返回 / four tabs 总览 · 策略 · 战斗累计 · 对局记录 /
导出 导入 清空, a note line saying the data stays in this browser and that teammates' names are stored too), text sizes
with max(.xxrem, N px) for phones in landscape, a 50-row history page with 显示更多, a pass-rate bar, a per-difficulty
table in order; the entries use a new `chart` icon (not the ping bars), the title screen's sits beside the language menu
(the connection row has no room: it squeezed the gear and wrapped the status text), the lobby / room buttons show the
icon alone below 700 px (the room's title ran under the label at 640 px). docs/PLAYING.md §12, README row; 41 + 25 UI
strings in en / ja / ko / zh-TW (the PR's four unused ones removed).
Tests: test/ui/stats.test.js 44 / 44 (counting rule per phase, quit records, bounds, corrupt / old / newer storage,
import order, aggregation, the replay payload); test/ui/stats.e2e.test.js (opt-in SP_E2E=1, real server, headless Chrome:
seeded history on 1920×1080 and 844×390, the re-view, export / import / clear, corrupt and newer storage, and the real
flows: a solo match given up in the briefing is history only, one given up after round 1 counts, a settled match is
recorded once and not again after a reload) 6 / 6; lint 0 errors, i18n extract 0 / check --all --strict clean,
check:imports, typecheck, docs and version tests, golden unchanged.
- Conflict: public/js/ui/settings.js imports (store from #321 + VOICE_LANGS) — both kept. test/ui/playtest6-elite.test.js's
source pin of the onDetail line allows the new tap counter (bcee737f).
- Grok review: MERGE AFTER FIXES — the three findings fixed on the branch (bcee737f). Golden unchanged; broad node --test
5334 (the one failure was that source pin, fixed), i18n strict clean, lint 0 errors, typecheck.
1. VoiceGate.start no longer moves lastAt for VOICE_TAP_SLOTS: a tap still skips the 1.2 s gap but starts none, so a
部署 / 技能 / 接敌 the battle asks for once after a short 选中 line has ended plays where the last battle line's gap
ends (a 部署 at 0, a tap at 500: a 作战中 at 1300 was dropped, now plays; inside the 部署 gap it still waits).
2. Shop / reward cards hand only the chess id to the detail, so two cards of one operator (the pool deals duplicates)
had one 选中 key and the second tap said nothing nor replaced the first line: game.js numbers every card it opens
(`tap`, cardTap), resolveDetail keeps it, and detailPanel's new selectVoiceKey() includes it (a re-render of the
same tap keeps its key; pieces / units unchanged). Checked in the mock match: the second card of 歌蕾蒂娅 now calls
voice() (it made no call on the parent).
3. tools/package.mjs FULL_ZIP_JP_VOICE comment: 89,388,041 bytes = 85.2 MiB on disk, about 76 MB zipped (as ASSETS /
DEPLOY), not "~93 MB, barely compresses".
Tests: test/ui/audio.test.js (the tap-gap case; one earlier assertion that encoded the old gap after a tap replaced),
test/feedback7-voices.test.js (new: the card tap in resolveDetail / selectVoiceKey / game.js) — both fail on the parent.
§21.30's 0.2.2 note updated.
Applied 设置 → 问题反馈. The report stays in the browser: the in-memory error log, device info, and the battle on screen only while 「附上本场战斗」 is on. Nothing is uploaded.
Redaction no longer treats a long name as a bare substring. Prose replaces each secret only as a whole token, so a name does not cut a longer word or a chess id. The attached JSON replaces a string only when it equals the secret, so chess_char_1013_chen2 and a note like "charge the char" stay, while an exact "char" value becomes the placeholder. The token boundary is a consumed prefix plus a lookahead, because a lookbehind is a SyntaxError in Safari before 16.4. PLAYING section 12 stays; the other section-12 proposal is not in this batch.
Tests: node --test test/diag.test.js — 10 pass. node --test test/client-static.test.js — 315 pass. node --test test/match/runner-pending.test.js test/match/watch-bonds.test.js test/ui/playtest6-unite.test.js test/ui/spectate-effects.test.js test/ui/feedback5-hotkeys.test.js test/i18n-server.test.js test/ui/bosshit-ticker.test.js test/ui/playtest3.test.js test/docs-consistency.test.js test/docs-paths.test.js test/version.test.js test/package.test.js — 138 pass. npm run -s lint — 0 errors. npm run -s typecheck and npm run -s check:imports passed. node tools/i18n.mjs check --all --strict — 1041/1041 in en, ja, ko and zh-TW, 0 errors.
- public/js/ui/stats.js: one record per finished match in localStorage (sp.pref.stats, envelope {v, records}, at most
1000, quota shrink of the oldest), content-id dedupe of the lobby's result replay, a MIGRATIONS chain and
normalizeRecord (defaults for missing fields, unknown fields kept, a newer envelope refused), merge-by-id export /
import, aggregateStats.
- public/js/screens/stats.js: the viewer (总览 / 策略 / 称号 / 战斗累计 / 最近对局, a row re-views that match's
settlement), 导出 / 导入 / 清空.
- Applied from the PR's own diff after reading it in full (no network, no new dependency, server untouched); only the
UI strings of the page are added to the four packs (en from the PR; ja / ko / zh-TW translated for this commit).
- The follow-up commits rework the page (the counting rule for matches left early, storage bounds, the layout).
- Tests: test/ui/stats.test.js 20 / 20; i18n extract 0, check --all --strict clean.
- With HTTP(S)_PROXY set, Node 22.0–22.20 / 23 cannot route the built-in fetch through it; the port of #289 made
fetch-assets (and so setup) fail there. Before 0.2.2 those downloads went direct, and a proxy variable left over
from another tool must not stop a player's first start: restartForEnvProxy now logs the reason and returns false
(go on in this process), guardDefaultFetch warns once and fetches. A Node that can proxy (≥ 22.21 / ≥ 24) still
restarts once with NODE_USE_ENV_PROXY=1 and never fetches around the proxy.
- restartForEnvProxy takes an injectable `version` for the test. Tests: assets-env-proxy 9 / 9 (a new case for the
old-Node path).
Applied the PHASE.PREP double-click on a teammate's avatar. 「前往查看」 stays. Boss-prep phases still use that button.
A teammate with no field is not opened: the click still toasts, and the double-click does not call onWatch. A click while that teammate is already on screen does not open the view again, including the second click of a double-click in server-run combat, where the first click already watches.
Tests: node --test test/ui/prep-dblclick.test.js test/ui/playtest6-unite.test.js — 13 pass. node --test test/ui/playtest3.test.js — 25 pass (same run, before the source-assertion fix; that file was not changed again). npx eslint public/js/ui/teamPanel.js test/ui/prep-dblclick.test.js — 0 errors.
Applied the host-only room option 「AI 队友最后选择」, default off. Off, the draft order and the random streams stay as they are. On, humans draft before AI seats after the existing shuffle, and a skip stays ahead of the AI seats.
Dropped docs/history/0.2.2.md and every §27 citation. The player-facing sentence is in docs/PLAYING.md. The room-option rule is in docs/design/match.md, docs/design/client.md and docs/design/network.md. The #338 per-bot faction and ban presets are not included.
Tests: node --test test/lobby-ai-last.test.js test/match/ai-picks-last.test.js test/ui/ai-picks-last.test.js — 18 pass. node --test test/lobby.test.js test/client-static.test.js — 378 pass. node tools/i18n.mjs check --all --strict — 1032/1032 in en, ja, ko and zh-TW, 0 errors.
Seen (reproduced on the parent, notes repro-354.mjs): band_orchid's active refresh copies the top-tier operator into another
slot and freezes the copy (bands/meta.js band_shop_refresh_copy_max_lv_char, slot.frozen = true) while shop.frozen — the
toggle — stays false; m.private sends both (views.js: shop.frozen and slots[i].frozen). The shop bar passed the toggle
(`frozen = !!shop.frozen`) to every ChessCard / ItemCard, so no card showed the frost: the server kept the card to the next
prep, the player could not see which one.
Now: shopBar.js slotFrozen(slot, toggle) = the slot's own flag, or the toggle (a frame without per-slot flags — the dev
mock — and the old behaviour for the 冻结 button, which still copies the toggle onto every unsold slot); both card kinds
use it. The bar's frame (is-frozen) and the 冻结 / 解冻 button stay on the toggle. The rule is not touched (bands.test.js,
and the new test): a second manual refresh re-rolls the frozen card too — like the toggle, "再手动刷新会重抽" — and
freezes one of the new pair; with no further refresh the frozen card stays in the next prep (the flags are cleared
then, round.js startRound).
- Tests: test/ui/feedback7-orchid-freeze.test.js (4: the wire — exactly one frozen card of the pair, toggle off — and the
cards drawn from it; the toggle still freezes every slot; the rule unchanged across two refreshes and a round; the
ShopBar source pin); fails on the parent (no slotFrozen, the cards passed `frozen`).
- Docs: PLAYING.md (冻结 row), design/network.md (m.private slot.frozen). Golden: unchanged (client render only).
Applied unchanged after review (Grok static review: MERGE; the diff read in full: no network, no dependency).
UnitView._updateImpostor refreshed a unit when (frame + phase) % interval === 0 with a random phase per unit, so the
refreshes of a frame were binomial: reproduced on the parent with 7 pen figures at interval 3 — 5 1 1 5 1 1 … and
2 4 1 2 4 1 … refreshes per frame instead of 2 / 3. Now render/app.js hands every impostor unit the next slot of the frame
(ctx.impostorSlot, a counter reset in frameBody), the unit refreshes when (frame + slot) % interval === 0, and a unit
whose slot keeps moving with the frame (the units before it culled on and off in step) is still refreshed after 2
intervals (imp.last); a context without slots keeps the random phase. Each unit's rate and the total work are unchanged;
no sim or pixel-as-state change.
Step 2 of #327 (a time budget by priority) is not part of this and still needs a decision.
- Tests: test/render/unitview.test.js 'impostor refresh slots' (3: slots in update order — fails on the parent, the
starvation cap, the no-slots context); unit view / render suites unchanged.
- Docs: design/client.md. Golden: unchanged (client only).
Seen (Grok's review, reproduced here with the real sim): the sim's deploy emits a deployment-activated skill's
['skill', id, 1] before the unit's ['spawn', unitInfo] (宴: three events earlier) and audio.js dropped a skill event of an
unknown unit, so that one cast made no sound while every later one did.
- audio.js: pendingSkill — a 'skill' ON for an unknown unit is held (one per unit, 64 at most, the oldest go first);
_track plays it once when the unit's spawn arrives; a skill OFF is no cast and is not held; setFieldUnits drops the holds.
Changed from the PR: setFieldUnits clears the holds BEFORE it tracks the new field's units (the PR cleared after, so a hold
left over from the previous field would have sounded for a unit of the new field that shares its id — ids are per battle).
- screens/game.js: NOT the PR's change. The pre-entry buffer (keepEarly: every spawn, die, deploy, status and skill since the
field's m.field, up to 1500 of them) is handed to the sound as its 'spawn' tuples only (new early.js audioEarly): the whole
list would play every old death, deploy line and cast at once on entering. The live path (onEv) already forwards the full
list, so the hold alone fixes the live silent first cast; a deploy-tick cast that falls into the short pre-entry gap
stays unheard (a deliberate gap — replaying "only the casts before their spawn" would still sound a minute-old cast when a
field was buffered while another one was on screen).
- Tests: test/ui/audio.test.js (3: the held cast in one batch and across batches, once, no hold for a skill OFF; a hold does
not outlive its field and is bounded; 宴's real deploy: the sim's cast comes before the spawn and sounds exactly once — all
fail on the parent) and test/match/feedback1-runner-forms.test.js (the audio gets the spawn tuples only: silent replay that
still teaches the units, the next live cast sounds; the whole list would have sounded old cues). Docs: design/client.md.
- Golden: unchanged (client audio only).
Applied unchanged after review (Grok static review: MERGE; the diff read in full: no network, no new dependency, no
shipped eval). The leak was real on feedback7: mountFieldView raced createFieldView against a timeout and, on failure,
removed every child of the host while the factory kept running — Pixi finished later and appended its canvas over the
simplified view (taking its clicks); the host wipe could also delete a newer view mounted by the same hook.
- ui/fieldHost.js: mountFieldView(host, { signal }) owns an AbortController; the timeout, a failed import or an invalid
view aborts it, a view a factory returns after that is destroyed, an outer abort (useFieldView's cleanup) rejects
without a fallback; nothing clears the host's children any more.
- render/app.js createFieldView takes the signal: the Pixi load / manifest / font waits end on abort, the synchronous
build registers its disposables (setupCleanup) so a throw releases the app, tiles, fx, impostors, listeners and
observers it made, the board-art and 3D-board waits moved after the view exists (aborting them is view.destroy(),
idempotent), destroy() also removes the canvas. render/app/host.js withTimeout(p, ms, signal).
- Tests: test/ui/fieldHost-lifecycle.test.js (5, node vm, fake timers: all fail on the parent) and the opt-in real-browser
test/render/initialization.browser.test.js (6 scenarios in headless Chrome, 5 fail on the parent: a late Pixi load
covering the fallback — with a factory that honours the signal and one that ignores it —, art timeout, a setup
exception, an unmount racing a newer mount); test/ui/playtest2.test.js's source pin takes the new indentation.
fieldview / browser suites in Chrome unchanged (the two 4× throttled fps tests flake when two Chrome files run
at once, pass alone).
- The createFieldView body stays unindented inside its new try (the PR's choice, to keep the diff reviewable).
- Docs: design/client.md §9. Golden: unchanged (client lifecycle only).
The three PRs patched the same snapshot and HUD lines (Grok's review: do not merge three diffs), so this is one design
built from their code and tests; IceCodeNew is the commit author (#303, #378), Convey123 co-author (#383).
- b.snap gains three optional, display-only lists beside the unchanged 9-field tuples (snapshot.js ammoView / wolfView /
negView, called from battle/events.js snapshot(); a living, deployed, visible ally, only when it has one):
ammo [[id, rounds left, magazine]] (whole rounds; the magazine is the activation's real total that unitTuple's SP
fraction runs on, extras and refills included), wolves [[id, 狼影 left, talent maximum]] (mem.shadows of the new
mem.wolfCapacity, set by tokens.js wolfPack and the managed vigil kit), neg [[id, fill]] (unit.negFill(), a function
op-hsgma2.js sets: the pool is read when the snapshot is taken, no per-tick sync). UnitInfo.ammoSkill (an ally whose
skill is an ammo magazine) is known from the first appearance: fx events run before the frame's sample, so the aura
cannot wait for the first ammo row.
- render/interp.js keeps the lists per snapshot (malformed rows / unknown ids dropped) and sample() hands them out from the
older snapshot like flags (they step with the skill flag); render/units.js draws them: the ammo magazine replaces the
SP bar (a cell per round, thin cuts, exact fill, no cuts under 2 px a cell), a row of diamonds under the bars for the
wolves (the element gauge row moves below), the red bar over the drained HP bar for the pool. An ammo skill draws no
standing aura (rings.js _aura; the activation burst stays).
- PR #378's wolf-head avatars are NOT taken: the count is data (the talent text 至多3只, mem.shadows), the HUD art of the
official counter is in no game data we have, so the pips are plain diamonds [ASSUMED look]. The PRs' own tests are
folded in and extended (test/sim/hud-readouts.test.js, test/render/hud-readouts.test.js, op_hsgma2.test.js).
- Docs: network.md, SIM.md §9, client.md, PLAYING.md, types/core.js.
- Golden: unchanged (283 scenarios; the canon hashes units / down / elem only, no sim state moved).
- Tests: 4 + 14 new + 1 in op_hsgma2 (12), all fail on the parent; 344 related tests pass; lint 0 errors, typecheck,
check:imports, docs tests.
Co-authored-by: Convey123 <Convey123@users.noreply.github.com>
SpineActor.attack() starts an attack that was not wound up (no look-ahead: a batch that arrived late) on its strike
frame — right for a quick rhythm, but for a slow clip-per-attack enemy the wind-up was cut out: 重犯 (Attack_grey
2.333 s, strike at 1.0 s, attacks every 3 s) showed the slam without the ball ever rising, and only 1.33 s of the clip.
The owner's decision of 2026-10-07 on #246: re-implement on the current spine.js so that, when the attack interval
leaves room for the whole clip, it plays once completely from the start. A clip-per-attack actor (enemies) with
interval ≥ the clip now plays it from 0 at its own speed for the whole clip, the strike landing a little late; a
quicker rhythm keeps the strike-frame start, and operators' looping attacks are unchanged (the PR's condition had it
the other way round: it changed the looping attacks and left out the clip-per-attack enemies it was written for).
Tests: test/render/feedback2-enemy-attack-clip.test.js (+1: 重犯 plays the whole clip from its wind-up, a quicker rhythm
and an operator start on the strike frame) 5/5; render eyemask, unitview, texas skill clip, skill idle 51/51.
The eyelids of 佩佩, 仇白, 隐德来希, 琳琅诗怀雅 and other models are Spine clipping attachments: the polygon is the eye's
opening and the eyeball / eye-white meshes drawn after it are clipped to it, so a blink or a knock-down closes the eye
through the mask. render/app.js pickClipping keeps the masks only for a lone clipped skeleton at high quality, so in a
battle the eyeballs showed through the closed lids. SpineActor._eyeMaskFallback now runs after each update while the
masks are off: it walks the draw order like pixi-spine (from a slot whose attachment in force is a clip, to its
endSlot — any skin) and hides the slot container of a covered slot whose drawn centre (a mesh's mean vertex, a region's
quad centre) lies outside the clip polygon [ASSUMED: all or nothing per slot]; turning the masks on shows them again.
PR #384's version hid `slot.currentSprite`, which pixi-spine leaves null on a mesh slot — every covered slot of the
417 local operator skeletons' 73 clipped ones is a mesh (313) or a region (54; regions have no worldVerticesLength, so
it never measured them) — and its test stubbed a currentSprite onto every slot, so nothing was hidden in the game. Kept
from it: the diagnosis, the ray-casting pointInPolygon and the key-frame expectations.
Tests: test/render/eyemask.test.js (new, 7, on real runtime objects — the official skeletons in @pixi-spine/runtime-3.8's
real Spine, @pixi/core display objects with a 16 px canvas stand-in, the real SpineActor: the eyeballs are meshes in
the clip's container; 佩佩 / 仇白 Die and 隐德来希 / 琳琅诗怀雅 Idle hide exactly the eye slots in the closed frames and
nothing open; a 30 Hz sweep of six clips; masks on hide nothing, turning them on shows all) 7/7 — the five skeleton tests
skip without public/assets (not committed; npm run assets), and fail on PR #384's code; render unitview, enemy attack
clips, texas skill clip 45/45.
The owner reported 「琳琅诗怀雅的2技能的香槟会掉血，会被治疗」. Allies always draw an HP bar (render/units.js
_updateHud), so the bomb's every HP change showed. PRTS 香槟炸弹 (实体类型 装置, 1000 HP) says no HP bar is shown even
below full HP; nothing there makes it invulnerable or unhealable. Its bar is now hidden at any HP (NO_HP_BAR). Its HP,
the damage it takes (活性源石 hurts what stands on it — 70 true damage a second) and healing are left as they are: no
invulnerability added; whether it should be unhealable is the owner's decision (no written source says so).
Tests: test/render/champagne-hpbar.test.js (new, 2: no bar at full HP, hurt or healed; another summon keeps its bar)
2/2; render/unitview 35/35.
- tools/assets/plan.mjs: one voiceTree() per dub — audio.voice (--voice-lang, cn by default) and audio.voiceJp
(VOICE_JP_LANG, always the JP dub, audio/voice/jp/<charId>/cn_nnn.mp3); fetch-assets plans and downloads both
(stats.voiceJpChars), so setup (checkAssets walks every listed URL) fetches both too.
- data/assets.json regenerated with the project's tool (node tools/fetch-assets.mjs --add-only, raw.githubusercontent.com,
107 s): the only change is audio.voiceJp + stats (+2674 files, +89,388,041 bytes, voiceJpChars 191) + hash — the rest
is byte-identical; the JP tree is the Chinese one with the folder swapped (191 × 14 files, every file a valid MP3).
The audio files are not committed (public/assets/ is git-ignored).
- public/js/audio.js voiceLine(): 'jp' draws from voiceJp and falls back to the Chinese line per slot (a JP entry the
manifest lacks) and per line at play time (the Chinese file of the same name when the host lacks the JP one —
same token, same gate); operators without voice (stand-ins, 盟约·辅助干员, summons) stay silent. setVoiceLang().
- settings: voiceLang 'cn' | 'jp' (VOICE_LANGS, sanitizeSettings), the 语音语言 row under 干员语音 (each dub named in
its own language, like the language switch), 语音语言 in the en / ja / ko / zh-TW packs.
- tools/package.mjs FULL_ZIP_JP_VOICE (true): false holds the JP files out of the full zip (and its update zip) —
not missing, not orphans, never in an update's `removed`; setup then downloads them.
- Sizes measured: art download 548 MiB (was 463), full zip ≈ 505 MB (0.2.1: 428 MiB + 76 MiB of zipped JP voice),
unzipped 711 MB (package --dry-run); lite zip unchanged. README, DEPLOY, ASSETS, setup's fallback size updated;
PLAYING voice paragraph; §21.30 "Changed in 0.2.2" note.
Tests: test/feedback7-voices.test.js (new: plan, manifest, settings, package switch), test/ui/audio.test.js (voiceLine,
the manager's dub choice and fallback), test/ui/gameLogic.test.js, test/docs-consistency.test.js — all fail on the parent.
The sim displaces at once (battle/displacement.js), so the snapshots only show the destination: a ≤ 2.5-tile push was
lerped over one snapshot interval (3 frames, 0.58 tiles in one frame — frame-loop repro over a real battle), a longer
pull snapped. render/app.js now starts UnitView.slideTo as soon as the interval it samples ends on the snapshot that
carries the destination (a `displace` fx still queued up to SnapshotBuffer.nextSnapT); the view holds where it stood
until that snapshot is shown, then eases into the sampled position under a constant deceleration (v0 = 2D/T, a = v0/T,
T = 0.14·√D in 0.12–0.45 s — ours, [ASSUMED]: no source gives the official timing; docs/research/13 has the official
impulse-and-friction classes). A unit that dies on the way finishes the slide and lies where the push put it.
Death: DIE_CLIP_MAX 1.6 → 8 s (data/assets.json: 234 enemy Die clips, the longest 7.97 s, 25 over 1.6 s), and a
flyer holds FLY_HOVER until the fade tail (DIE_FADE_TAIL 0.55 s).
Substantially @xcdoge's code (PR #380: the slide, the death timing, research 13 / 14, most tests), authored as theirs.
Changed on review: the corpse no longer goes back to where it stood (the PR's test asserted that; the sim's body is at
the destination); the PR's pre-pass scanned (renderT − 0.5, renderT] after processEvents had taken those events, so
it never fired and the slide started a frame before landing — it now looks ahead to the next snapshot and holds the
view until then; the slide eases into the live sample instead of the fixed destination (no jump back onto an enemy
that walked on); each step uses its mean velocity (the end-of-step speed ran 9 % short, then jumped); the constants'
comments say the timing is ours; research 14 no longer cites a sweep script and a data guard that were not in the PR.
Golden: unchanged (client only).
Tests: test/render/unitview.test.js 35/35 (slide: deceleration, idempotence, facing, snapshots in flight, friction,
√D, corpse lies where pushed, hold-then-ease without a landing jump, a never-shown slide handed back; the app.js frame
order over a real battle: a 1.7-tile push over ≥ 8 frames with every step < 0.35 tiles, a push-kill's body staying at
the destination; death: a dying flyer's lift, every Die clip within DIE_CLIP_MAX and a 7.97 s clip scheduled whole);
test/sim/displace-fx.test.js 1/1; render/interp 16/16; docs tests pass.
Did not apply the patch. It set a global dispatcher on the npm undici package and rewrote the lockfile. Node's built-in fetch does not use that dispatcher, and node:undici is not importable. fetch() reads HTTP_PROXY, HTTPS_PROXY and NO_PROXY only at startup, and only when that startup had NODE_USE_ENV_PROXY=1 (or --use-env-proxy). That is Node >=22.21.0 and >=24.0.0.
tools/fetch-assets.mjs restarts itself once with NODE_USE_ENV_PROXY=1 before downloading. The default fetch in the downloader and the index cache rejects when a proxy is set but this process cannot use it, including Node 22.0–22.20 and 23, instead of fetching directly. No new dependency. Engines stay >=22.
Tests: node --test test/assets-env-proxy.test.js — 8 pass, 0 fail. node --test test/asset-network.test.js test/assets-diy.test.js test/fetch-assets-shrink.test.js test/assets.test.js — 84 pass, 0 fail. Before this commit: npm run -s lint — 0 errors (135 existing warnings); npm run -s typecheck — ok; npm run -s check:imports — exit 0 (3 pre-existing sim-node notes). docs, version and package tests: 46 pass and 1 fail while env-proxy.mjs was still untracked; after git add, node --test test/package.test.js — 10 pass, 0 fail.
Since 0.2.0 the sim stands an unblocked enemy for the rest of its attack clip (ai.js attackStand → atkStandUntil),
but render/interp.js lerped x/y over the whole snapshot gap, so the view slid through a recovery the sim held
(repro: a real 隐形弩手, snapshots 4.9 / 5.0, stand until 4.934 — sampled at 4.917 it slid at vx −0.2).
The snapshot now carries, only when non-empty, `stand: [[id, until]]` (atkStandUntil of enemies alive, deployed,
visible, neither feared nor stunned) and `standCut: [[id, at]]` (the latest time each listed enemy's recovery was cut
or ignored: a displacement, fear, a stun, hiding, leaving — Battle._cutAttackStand, display metadata the battle never
reads). sample() holds the older position until `until` and interpolates the rest (HP / SP keep the whole interval),
stays linear when the newer snapshot records a cut inside the interval, and extrapolates nothing for a newest snapshot
that shows a stand, a death, a stun or a block. Teleports and redeploys still snap first; old snapshots without the
fields interpolate as before.
Substantially @ZiyueW1116's code (PR #381), authored as theirs. Changed on review: the docs (SIM §9, design/network.md)
and code comments in English like the rest of them; the snapshot builds its listed-id set only when an enemy has a cut
to report.
Golden: unchanged — the digest canon hashes neither field and the sim's timing does not move (all 283 scenarios match).
Tests: test/render/attack-recovery-interp.test.js (new, 25: metadata validation, hold-then-slide, cuts, legacy
snapshots, extrapolation guards, real 隐形弩手 / Weedy push-kill / stun / freeze / sleep / levitate / block / fear
battles) 25/25; render/interp 16/16; sim battle, perf, robustness, downelem 69/69.
screens/game.js passes the detail panel voice=${true} instead of its combat flag, so the panel's existing
audio.voice(charId, 'select') runs on every opened operator: a piece on the field or in the hand, a shop /
reward card, a bond member. The gate follows the official FOCUS_CHAR row (audio_data.json battleVoice:
priority 10, cooldown 0, overlapIfSamePriority true): VOICE_TAP_SLOTS — on an idle channel a tap always
plays (no 1.2 s global gap; VOICE_COOLDOWN_MS.select 1500 → 0), a newer tap replaces the 选中 line on air,
and a higher-priority line (部署 20, 作战中 70, 开战 100 …) is never interrupted. Deploy / skill / engage /
result lines stay battle-only; stand-ins, 盟约·辅助干员 and summons stay silent (no voice entry).
Tests: test/ui/audio.test.js (new: the gate and two taps through the manager; fails on the parent),
test/docs-consistency.test.js pins updated. Docs: §21.30 "Changed in 0.2.2" note, PLAYING.md voice paragraph.
Ported only the dual-stack listen from PR #291 (the change accepted as #188). With HOST unset the server binds :: and, if that fails with EAFNOSUPPORT, EADDRNOTAVAIL or EINVAL, retries 0.0.0.0. An explicit HOST, including 0.0.0.0 and 127.0.0.1, is tried once. The boot banner and doctor list shareable IPv6 addresses in brackets, skip fe80::, and keep one URL per /64. An explicit IPv6 literal in the local URL is bracketed. The Windows service listing no longer drops those bracketed lines.
Not ported: /dev/grant, the settings-default migration, and the firewall scripts. The Docker image stays on 0.0.0.0; the alpine fallback was not booted.
Tests: node --test test/ipv6-bind.test.js test/doctor.test.js test/ci.test.js — 20 pass, 0 fail (8 bind, 4 doctor, 8 ci).
The 赤霄·天喟 wave is a kit-managed probe (it turns corners and lives as long as the skill), not a snapshot
projectile, and the sim emitted one `dash` at the cast and a `slash` per hit — nothing along the way, so the client
drew no wave. op-chen3.js now emits fx('chen3Wave', { x, y, fromX, fromY, dr, dc, id }) every WAVE_FX_EVERY (0.12 s;
4 ticks), the trail restarting at a corner; render/fx registers chen3Wave → 'qi' (pt: drawn at the event, never on
her) and draws a streak plus a crescent head turned the way it travels. The look is a procedural stand-in, not the
official particle prefab.
Substantially @Convey123's code (PR #382), authored as theirs. Changed on review: the head is projected after
streak() — streak() reuses the effect system's projection scratch, so the PR drew the crescent on the trail's start,
one event behind the wave — and the trail, head and heading share one height; an unused test variable dropped.
Golden: diy-073 (5_diy*=chen3 S3 in two slots) events.fx 508 → 1558, fx.chen3Wave → 1050; diy-074 (6_diy*=chen3 S3)
events.fx 272 → 872, fx.chen3Wave → 600. HP, kills, damage, hooks and snapshot hashes unchanged: battle.fx only
records an event.
Tests: test/content/op_chen3.test.js (+1: the events, their cadence, id, direction, trail start, silence after the
skill) 12/12; test/render/fxchen3wave.test.js (new, 6: the archetype, drawn at the wave not at her, the head on the
wave and turned with it, a sweep joining into one trail, junk input, nothing left after) 6/6; boardart 15/15.
Applied as-is. Captured real battles live in public/dev/perf (including the round-13 联防 on the wide round battlefield), with a browser speed page and a headless benchmark. public/dev stays out of the package. The Chrome benchmark was not run. The numbers in the original pull request are from 0.1.4 and are not quoted here.
Tests: node --test test/perfbench.test.js test/render/battle-perf.browser.test.js — 13 pass, 0 fail; the Chrome suite skipped (RENDER_E2E unset).
Applied as-is. tools/ci.mjs runs the checks already listed in .github/workflows/ci.yml, in that order, and skips checkout, setup-node and npm ci. Children are spawned with process.execPath, not a shell. test/ci.test.js keeps the step list, smoke script, job env and package.json scripts in step with the workflow.
Tests: node --test test/ci.test.js — 8 pass, 0 fail. npm run ci was not run (it installs and runs the suite).
Applied as-is. All-mode verification runs the existing HeadlessJob in headlessSliceMs slices (8 ms on a real scheduler; synchronous when that value is not finite, including the virtual scheduler). The field still settles with the server digest. Sample mode still accepts the client result immediately and logs afterwards. A comparison error in all mode falls back to the server result. Dispose cancels the pending slice timer.
Tests: node --test test/match/verification-slices.test.js — 11 pass, 0 fail; node --test test/match/clientCombat.test.js test/match/clientCombat-review.test.js — 39 pass, 0 fail.
Applied as-is. The client precheck for withdrawing a board operator onto an occupied hand slot now treats that operator's own hand summon stack as a free slot, which the server already accepted. Temp stacks, board summons, another player's tokens, and an item that only carries ownerUid stay HAND_FULL. stack.count is not extra slots. Dropping on a chess still checks the swap first.
Tests: node --test test/ui/gameLogic.test.js test/ui/facing.test.js test/ui/uiFixes.test.js test/match/feedback1-placement.test.js test/match/feedback5-we2-placement.test.js — 116 pass, 0 fail. Before this commit: npm run -s lint — 0 errors (134 existing warnings); npm run -s typecheck — pass; npm run -s check:imports — pass; node --test test/docs-consistency.test.js test/docs-paths.test.js test/version.test.js test/package.test.js — 47 pass, 0 fail. No new UI strings, so the i18n check was not run. No golden update.
Applied as-is. The replace dialog no longer treats Enter that started on a button as confirm. Enter or Space on 取消 closes the dialog and leaves the chosen equipment in place. Enter on an option only changes the selection. Enter on the confirm button still sends g.equip with that replaceUid. This matches the guard already used by DialogHost.
Tests: node --test test/ui/mock.e2e.test.js test/ui/leftovers.test.js test/ui/standin-ui.test.js — 31 pass, 0 fail. The mock e2e suite skipped because SP_E2E is unset. eslint on the touched files: 0 errors (one pre-existing unused-assignment warning in mock.e2e.test.js).
Applied as-is. The shop bar's two-tap arm for the level button is now lv:<level>. A successful upgrade, including the D shortcut, changes the level and drops that arm, so the next click only selects the new level. A freeze at the same level keeps the confirm. Another card's arm is left alone, and D still upgrades immediately.
Tests: node --test test/ui/mock.e2e.test.js test/ui/playtest6-elite.test.js test/ui/playtest6-ui.test.js — 20 pass, 0 fail. The mock e2e suite skipped because SP_E2E is unset. eslint on the touched files: 0 errors (one pre-existing unused-assignment warning in mock.e2e.test.js).
Applied as-is. Prep-scout refresh now signs the public prepFieldMeta payload, so a 机变 or bounty change that only updates effects or nextEnemies is pushed to teammates and spectators. Shop, funds, and private views stay off that payload. An unchanged public view is still not sent twice, and an explicit g.watch still resends the current view.
Tests: node --test test/match/prep-watch.test.js test/match/prep-bench.test.js test/match/spectator.test.js test/match/feedback5-boss-prep-scout.test.js test/match/teammate-loadout.test.js test/ui/spectate-effects.test.js test/match/feedback5-boss-pair-prep.test.js — 25 pass, 0 fail. eslint on the touched files: 0 errors.
Applied as-is. BUILD_INPUTS now covers server/sim, shared and data as well as the page, so a restart after a simulation-only or data-only change is visible to an already-open page. The tag is still computed once per process. A page still waits until the match, including settlement, is over before it reloads, and a failed check still does not reload.
Tests: node --test test/build.test.js test/ui/build-runtime.browser.test.js — 8 pass, 1 skipped (the Chrome restart test; SP_E2E unset), 0 fail.
sganggs