feedback5: merge feedback4 (beb0296) — PR #73 (operator battle voices) ported to the split layout (the voice setting into ui/gameLogic/settings.js, the gameLogic.js shim unchanged; onResult and DetailPanel voice stay in screens/game.js, which needed its store import back; the sim's 'engage' in server/sim/ai.js performAttack, still the one battle/combat.js and devices.js call; plan.mjs / fetch-assets.mjs merged with --add-only, extraOperators and moduleTypes; the 干员语音 slider through t() with English); golden: roster / bonds / fields / standins / diy differ only in events.engage, matches unchanged

Ported by hand:
- public/js/ui/gameLogic.js is a re-export shim on feedback5: #73's DEFAULT_SETTINGS / sanitizeSettings `voice`
  hunk went into public/js/ui/gameLogic/settings.js.
- public/js/screens/game.js: MatchScreen still lives there, so #73's onResult (battleRunner 'result' → resultSpeaker /
  resultVoiceSlot) and `voice=${combat}` on the DetailPanel auto-merged in place. feedback5's game.js no longer
  imported `store` (the split moved its other users), so onResult's store.get() would have thrown into its
  try/catch and the 结算 line would never have played: `store` is imported again. `localAsset` stays dropped.
- public/js/ui/settings.js (i18n phase 1): the slider is `label=${t('干员语音')}`, public/i18n/en.json
  "干员语音": "Operator Voices" (next to Music / Sound Effects).
- tools/assets/plan.mjs: buildPlan takes #73's charword / voiceLang / voiceSlots and feedback5's extraOperators /
  moduleTypes; the template has `modules` and audio { bgm, bossBgm, voice, sfx }. `chars` now includes the 自选 picks,
  so the voice loop's comment says so (the next `npm run assets` plans their lines too).
- tools/fetch-assets.mjs: parseArgs keeps --add-only beside --voice-lang / --voice-all.
- data/assets.json: feedback5's manifest plus #73's audio.voice (120 operators, after bossBgm, the plan order);
  stats files +1680 and bytes +42 041 118 as measured on feedback4, voiceChars 120; the hash is recomputed with
  the generator's contentHash. No other entry changed.
- docs/ASSETS.md: the stats line lists modules and voiceChars. DESIGN §21.30 stays where feedback4 put it; SIM.md's
  'engage' line and docs-consistency's §21.30 test auto-merged (game.js / detailPanel.js still hold the pinned code).
- Auto-merged and checked: audio.js, battle/runner.js emit('result'), components.js mic icon, detailPanel.js voice
  prop, shared/protocol.js EV.ENGAGE, server/sim/ai.js, tools/assets/audio.mjs / cache.mjs, #73's tests.

Where #73 met feedback5 work:
- 补位: a chess fielded as its stand-in shows the stand-in in the detail card, and audio.js keys battle lines by
  UnitInfo spine (the stand-in's model). The 选中干员 line now takes detail.standIn.charId first, so the stand-in
  speaks, not the operator it replaces (test/ui/standin-ui.test.js pins it). A 自选 record is already the pick's.
- The sim's 'engage': stand-in and 自选 units attack through the same performAttack, so it fires for them as well.

Golden regenerated: every moved scenario differs only in events.engage (the new event's count; no gameplay change),
proven by a field-path diff against feedback5:
- roster: all 49; bonds: all 46; fields: all 22; standins: all 10 (the stand-ins engage).
- diy: 101 of 109. Unmoved: diy-010/011/012/092/093 (闪灵, 夜莺, 凯尔希 squads: they only heal, and a heal never
  engages) and diy-077 (薇薇安娜), diy-080/081 (赫德雷), whose picks make no attack in those battles (already so on
  feedback5).
- matches: unchanged.

docs/ASSETS.md

@@ -19,7 +19,9 @@ npm run assets       # = node tools/vendor.mjs && node tools/fetch-assets.mjs
 | `--force` | Re-download everything. |
 | `--offline` | No network. Re-runs post-processing (atlas fixes, skeleton parsing, WOFF2) on what is already on disk, then rebuilds `data/assets.json`. |
 | `--dry-run` | Print the plan (file and model counts, alias notes) and exit. |
-| `--refresh-index` | Re-download the two upstream indexes: `audio_data.json` and `models_data.json`. |
+| `--refresh-index` | Re-download the upstream indexes: `audio_data.json`, `charword_table.json` (the 干员战斗语音 slots) and `models_data.json`. |
+| `--voice-lang=cn` | 干员战斗语音 language: `cn` (default) | `jp` | `en` | `kr` — the same file names under `voice_cn/`, `voice/`, `voice_en/`, `voice_kr/`. |
+| `--voice-all` | Plan every official voice slot, including the prep-only lines no battle plays (干员报到 / 编入队伍 / 任命队长 — 360 files, one per operator and slot). Off by default: nothing requests them, so planning them only makes every run download more. |
 | `--prune` | Delete files under `public/assets/` that the manifest no longer references, for example after a mapping change. Without this flag they are only listed in the report. `public/assets/local/` (written by `tools/local-extract`) is never pruned. Implies `--allow-shrink`. |
 | `--allow-shrink` | Write `data/assets.json` even when it loses entries the current one has (see "The manifest never shrinks by accident" below). |
 | `--add-only` | For a checkout whose `public/assets/` and `public/fonts/` are shared with another one (a git worktree with symlinked asset folders): download only the files missing on disk and never re-download, rewrite or delete an existing file — atlases already on disk are left as they are, the fonts are not rebuilt (the manifest keeps its current `fonts`). Not with `--prune` / `--force`. |
@@ -32,7 +34,9 @@ audio entries short — all 42 still resolve upstream; PR #7). When the rebuilt
 one, the run keeps the current file, prints the entries it would drop (also in the report: `droppedEntries`,
 `manifestWritten: false`) and exits 1. Re-run to retry the downloads, or pass `--allow-shrink` (or `--prune`) when the
 smaller manifest is intended, for example after a mapping change. Build fields (`version`, `hash`, `generator`,
-`stats`), new entries and a changed value are never a drop (`tools/assets/manifest.mjs droppedEntries`).
+`stats`), new entries and a changed value are never a drop (`tools/assets/manifest.mjs droppedEntries`). A run whose
+plan legitimately narrows — like the 干员战斗语音 default, which no longer plans the three prep-only slots (360 entries,
+DESIGN §21.30) — reports exactly those entries and needs `--allow-shrink` once; the list it prints is the check.
 
 The script is **idempotent**. A file on disk is kept, not re-downloaded, when any one of these holds:
 - its size matches the ledger entry from a previous download (`.cache/assets-ledger.json`);
@@ -49,7 +53,7 @@ How downloads are fetched:
 - A manifest entry with fallbacks (for example an enemy icon that falls back to its base enemy's icon) only moves on to the next alternative after a **definitive 404**. When the primary fails transiently (network error, 5xx or an invalid payload after all retries), no fallback is fetched. The path is listed under `downloadErrors` in the report, and the next run retries the primary.
 - A skeleton that fails to parse is deleted and removed from the ledger, so the next online run downloads it again.
 
-The first run downloads about **269 MiB in about 4,000 files** (it took 134 s on a ~3 MB/s link before the 55 emote and 玩法说明 files, 21.3 MiB, were added). A re-run takes about 1 s.
+The first run downloads about **309 MiB in about 5,690 files** (it took 134 s on a ~3 MB/s link before the 55 emote and 玩法说明 files, 21.3 MiB, and the 1,680 干员战斗语音 files, 40.1 MiB, were added). A re-run takes about 1 s. The voice count is the twelve slots a battle plays; the three prep-only slots the official client uses elsewhere (干员报到 / 编入队伍 / 任命队长, 360 more files, 18.4 MiB) are left out unless `--voice-all` is passed.
 
 Outputs:
 - `data/assets.json`: the manifest (committed).
@@ -88,6 +92,7 @@ The research JSONs in `docs/research/` (03, 05, 07) define **which** ids are nee
 | Enemy Spine that no dump carries (灼热源石虫 / 炽焰源石虫) | the local client only (`tools/local-extract/extract.py ENEMY_SPINES`, optional); never downloaded and never required: an overlay of the web alias (`enemies[id].spineLocal`) | `local/spine/enemy/{enemyId}/{stem}.*` (listed in `data/local-assets.json`) |
 | BGM | AA2 `voice` branch `audio/sound_beta_2/music/**` (大厅/休整期 `act1autochess`, 开战 `act13side/m_bat_kazimierz2_{1,2}` — 骑士之日 / 无畏者; the 开战 track follows the round: `_2` 无畏者 rounds 1–7, `_1` 骑士之日 from round 8) | `audio/bgm/{file}.mp3` |
 | SFX (UI, battle, per unit) | AA2 `voice` `audio/sound_beta_2/**`, mapped from `audio_data.json` banks | `audio/sfx/{same sub-path}.mp3` |
+| 干员战斗语音 | AA2 `voice` `audio/sound_beta_2/voice_cn/{charId}/cn_nn.mp3` — the lines `charword_table.json` lists (`placeType` = when the game plays one, `voiceAsset` = the path); `--voice-lang=jp|en|kr` takes the same file names from `voice/`, `voice_en/`, `voice_kr/` | `audio/voice/{lang}/{charId}/{cn_nn}.mp3` |
 | Fonts: Bender Regular and Light, Novecento Wide | TimWangZi/The-font-of-Arknights | `public/fonts/*.{otf,ttf,woff2}`, `public/fonts/fonts.css` |
 
 The `stem` of a Spine model is the upstream file name. Two examples: `char_107_liskam` has the stem `char_107_liskarm`, and `enemy_9032_aclionk` uses `enemy_1559_vtlionk`. The skel and atlas of a model always share one stem. pixi-spine locates the atlas by swapping the extension, so this matters.
@@ -138,7 +143,7 @@ All paths are URL paths relative to the site root, for example `/assets/char/ava
   hash: 'a1b2c3d4e5f6',             // content hash (cache busting)
   generator: 'tools/fetch-assets.mjs',
   stats: { files, bytes, chars, charsWithBack, enemies, enemiesWithSpine, tokens, tokensWithSpine,
-           spineModels, bonds, items, bands, skills, modules, ui, sfxUnits },
+           spineModels, bonds, items, bands, skills, modules, ui, sfxUnits, voiceChars },
   chars:   { [charId]: { avatar, avatarE2?, portrait, portraitE2?, spine: { front: Spine, back?: Spine } } },
   enemies: { [enemyId]: { icon, spine?: Spine, spineAliasOf?: enemyId,
                           spineLocal?: { group, skel, atlas, textures, …Spine } } },
@@ -165,6 +170,14 @@ All paths are URL paths relative to the site root, for example `/assets/char/ava
              // `bgmEvent = corrosion` (卡西米尔 act13d5d0), so the rescue phase does not reuse the 作战's track
              // (audio.js bgmKeyFor 'unite', falling back to `bgm.combat` for a manifest that lacks it)
     bossBgm: { [bossId]: { intro?, loop } },                   // per-boss track of its R14/R15 level
+    voice:   { [charId]: { start, faceEnemy, select, place, skill1…skill4,
+                           resultFour, resultThree, resultTwo, resultLose } },
+                           // 干员战斗语音: an operator's official battle lines (charword_table.json placeType → slot,
+                           // tools/assets/audio.mjs VOICE_SLOTS); a slot with several lines is an array and the client
+                           // draws one (public/js/audio.js voice). Only these twelve ever play (DESIGN §21.30): the
+                           // prep-only slots 干员报到 / 编入队伍 / 任命队长 are left out of the plan by default —
+                           // nothing requests them and they cost 360 files (19.3 MB) per run — and `--voice-all` adds
+                           // them (audio.mjs VOICE_PREP_SLOTS) for the complete official set
     sfx: {
       ui:     { click, back, confirm, tab, pick, drop, error, buy, sell, income, refresh, freeze, levelup,
                 merge, equip, itemMerge, bondUp, artPlace, ready, timer, draft, yourTurn, yourTurnCircle,

docs/DESIGN.md

@@ -1506,6 +1506,20 @@ Two public GitHub issues after 0.1.0, both client-only. Normative lines rewritte
 - Tests: `test/sim/feedback1-tank-triggers.test.js` (the data, kit = data, the real battle for the six and for every other MANUAL 重装 record), `test/data.test.js`, `kits_t1t2` / `kits_alt_t1` (PR #12's), `kits_alt_t2` (灰毫 S2), `kits_t5` / `kits_alt_t5` (号角), `test/docs-consistency.test.js`.
 - Extended 2026-10-04: 余 S2 (SKILL_RANGE on its x-1, §22.10); 14 records deviate now.
 
+### 21.30 干员战斗语音: an operator's official lines, in battle only (user request, 2026-10-03) — `tools/assets/plan.mjs` (`voiceSlots`), `tools/assets/audio.mjs` (`indexVoice` / `VOICE_SLOTS` / `VOICE_BATTLE_SLOTS` / `VOICE_PREP_SLOTS` / `VOICE_DIRS`), `tools/assets/cache.mjs`, `tools/fetch-assets.mjs --voice-lang / --voice-all`, `public/js/audio.js` (`VOICE_PRIORITY` / `VOICE_COOLDOWN_MS` / `VoiceGate` / `resultVoiceSlot` / `resultSpeaker` / `voice()`), `server/sim/ai.js performAttack 'engage'`, `battle/runner.js emit('result')`, `screens/game.js onResult`, `ui/detailPanel.js voice`, `ui/settings.js`, `ui/gameLogic.js sanitizeSettings`
+
+User: 「给每个卫戍干员都加语音」, then, after the first build, 「整备阶段不需要干员语音，战斗开始后才需要语音，也就是只需要作战中语音」.
+
+**Where the lines come from.** Every operator the mode can field already has official battle lines in the same voice dump the BGM and SFX come from — `voice_cn/{charId}/cn_nn.mp3` (JP/EN/KR: `voice/`, `voice_en/`, `voice_kr/`, the same file names, `tools/fetch-assets.mjs --voice-lang=jp|en|kr`). `charword_table.json`'s `placeType` says when each line is used, and `indexVoice()` turns it into the manifest's `audio.voice[charId]`: 行动出发 `start` (BATTLE_START), 行动开始 `faceEnemy` (BATTLE_FACE_ENEMY), 选中干员 `select`, 部署 `place`, 作战中1-4 `skill1..skill4` (BATTLE_SKILL_N), 编入队伍 `squad`, 任命队长 `squadFirst`, 结算 `resultFour/Three/Two/Lose` (完成高难行动 / 3星结束行动 / 非3星结束行动 / 行动失败) and 干员报到 `gacha`. A slot with several lines stays an array and the client draws one. Only the base word key is planned (`wordKey === charId`): the dump has no folder for a skin variant's word key (`char_x_ita`, `char_x_epoque#28`), so those lines 404. **Only the slots a battle can play are planned** (`audio.mjs VOICE_BATTLE_SLOTS`): 干员报到 / 编入队伍 / 任命队长 are lines of the official client's own 养成 / 编队 UI, nothing in this mode ever requests them (`test/docs-consistency.test.js` checks the client's `voice(...)` call sites), and downloading them cost 360 files (19.3 MB) on every run — `--voice-all` adds them back for the complete official set. Measured: 120 of the 121 in-mode charIds have a full set — the 17 预备干员 and the mode's own 盟约·辅助干员 have no battle voice at all, so they simply get no entry — 1680 lines of the 2040 upstream, in a manifest of 5694 files / 324.2 MB (files and `stats.voiceChars` in `data/assets.json`).
+
+**Where each line plays — battle only.** Every line comes from a running battle's own event stream or from its settlement, so no hook needs a phase check and the 休整期 stays silent: the shop does not say 干员报到 (a merge-completing buy neither says 任命队长), a retreat does not say 编入队伍 and dragging a piece onto the board in prep says nothing. In battle: 开战 — the first operator deployed says 行动出发 and the others 部署 (the sim's `deploy` events; `setFieldUnits` resets the flag per field — in this mode every piece enters at once, which is why the gate below exists, and a knocked-out operator's redeploy says 部署 again); 接敌 — the sim's new `['engage', id]` (a unit's first attack that hits an enemy, official ENCOUNTER_ENEMY); 技能 — `skillN` from the unit's `skillIndex` (作战中1-4); 选中 — the detail panel opening on an operator **while a battle runs** (`DetailPanel voice` = the game screen's `combat`); 结算 — `resultVoiceSlot()`: 完美作战 ⇒ 3星结束行动 (绝境 / 终极 ⇒ 完成高难行动), 有漏怪 ⇒ 非3星结束行动, 一个没杀 ⇒ 行动失败, spoken by a drawn operator **of that battle's own field** (`resultSpeaker`, from the result's `perPlayer.unitsEnd`: the operators of that field, survivors first, a summon piece never), driven by the runner's new `emit('result', …)` — never the field on screen, so watching a teammate cannot put their operator's voice on the viewer's settlement line (review on #73).
+
+**Why a channel, not SFX.** Voices are 1–4 s long and a battle deploys eight operators, then fires dozens of skills; playing them like SFX would flood the mix. `VoiceGate` copies the official scheduling data (`audio_data.json battleVoice.voiceTypeOptions`): one line at a time, a 1.2 s global gap, a per-unit per-slot cooldown (10 s for 作战中, 3 s for 接敌, 1.5 s for 选中) and takeover when a line's priority beats the playing one by a margin (the official 0.1 s `crossfade`) — the priorities are the official ones (BATTLE_START 100, ENCOUNTER_ENEMY 90, SKILL_ACTIVE 70, PLACE_CHAR 20, FOCUS_CHAR 10) and the settlement lines get 85, above 作战中 but below 接敌. Every deferred step of a line — its decode, its `onended`, the safety timer — carries its own token and returns without touching the gate once it is stale, so a line taken over while it was still loading can never free the channel its replacement just took (review on #73). The channel has its own gain and its own slider (设置 干员语音 / VOICE; `sanitizeSettings` adds `voice`, so a saved profile without it keeps the default and players can turn voices down without losing SFX). The buffer cache also gained a decoded-PCM budget (64 MB) beside its 180-entry cap: a voice decodes to ~0.4–1.3 MB, and 180 of them would be ~100 MB.
+
+**[ASSUMED].** 接敌 as "the unit's first attack on an enemy" (the official trigger is the first enemy it encounters, not necessarily one it hits) and the settlement mapping (漏怪 ⇒ 非3星, 完美作战 in 绝境 / 终极 ⇒ 完成高难行动) are our readings. Lines follow the field on screen: watching a teammate's battle (前往查看) plays *their* operators' lines, while the settlement line fires for the player's own battle only (`e.own`, `battle/runner.js`) and is spoken by an operator of that battle's own field (`resultSpeaker`). 部署 and 选中干员 stay in the plan although the official groups them with the prep lines: a battle *does* play them (the first operator says 行动出发 and the others 部署, and the detail panel says 选中 while a battle runs), so only the three slots no call site ever asks for are dropped — `VOICE_PRIORITY` / `VOICE_COOLDOWN_MS` keep every level, so `--voice-all` restores 编入队伍 / 任命队长 / 干员报到 for the day the 休整期 speaks again.
+
+- Tests: `test/ui/audio.test.js` (the priority table, the settlement slot and its speaker, the gate's global gap / per-unit cooldown / takeover, the stale-callback regression, and the battle events that drive the lines against a fake Web Audio), `test/ui/gameLogic.test.js` (the `voice` setting), `test/sim/battle.test.js` (the wire format: every emitted tuple is a known `EV` — the new `engage` included); the manifest's 1680 voice paths are covered by `test/assets.test.js`.
+
 ## 22. GitHub issues after 0.1.1 (v0.1.2)
 
 Public GitHub issues filed after the 0.1.1 release (2026-10-03), each triaged against 0.1.1 first; the real ones were fixed for 0.1.2 in fix branches WA–WL, each independently reviewed before it was merged into `feedback2`. Where each is handled: #25 back-facing operators keep their attack pose when knocked out → §22.1; #51 突袭's idle redeploy → §22.2; #33 items 1–2 阿戈尔's devour and 乌尔比安 S3 → §22.3; #33 item 6 活性源石, with the 沼泽 and 深水区 terrain → §22.4; #42 emotes and guide pages without a local client, and the English title (#38) → §22.5; #32 item 4 钩索师 / 推击手 on 高台 → §22.6; #52 史尔特尔's 余烬 → §22.7; #43 an enemy's 隐匿 after a block, and the 深池逐火 ember window → §22.8; #32 item 7 荒芜拉普兰德 S3's drones → §22.9; #32 item 1 余 S2 → §22.10; #44 归溟幽灵鲨's 替身 → §22.11; #32 item 6 enemy splash against a 隐匿 operator → §22.12; #35 / #16 (PR #48) 要塞 against air units and 莫斯提马's slow icon → §22.13; #58 / #60 / #64 follow in §22.14–§22.16. Decisions of the owner, 2026-10-04: 荒芜拉普兰德's drones measure the distance to an enemy's position, a huge enemy's centre (§22.9); 余 S2 casts with an enemy on its x-1 — a deliberate deviation from the official TAKE_DAMAGE row, added to §21.29's six (§22.10); 乌尔比安 keeps the devour gains through his 【移动】 — the deviation stays (§22.3).

docs/SIM.md

@@ -1254,7 +1254,8 @@ Unknown subprofessions fall back to the profession default (test `professions.te
   (`none|arrow|bolt|bomb|lob|orb|drone|enemy|boomerang|droneBomb|chain|chainHeal`; a boomerang's way back has no event — the
   renderer flies it back to the thrower at `BOOMERANG_RETURN_SPEED`; an enemy's `profile.shot` may name another kind,
   e.g. `mortar` for 帝国炮火先兆者, which the renderer does not draw — its fx `bombardShell` is the shell), `['dmg', tgt, amount, type]` (`phys|arts|true|burn|neural|necrosis|apoptosis`),
-  `['heal', tgt, amount]`, `['skill', id, 1|0]`, `['die', id, reason]`, `['leak', id]`, `['status', id, key, 1|0]`,
+  `['heal', tgt, amount]`, `['skill', id, 1|0]`, `['engage', id]` (an ally's first attack that hits an enemy — the client's
+  行动开始 voice, DESIGN §21.30), `['die', id, reason]`, `['leak', id]`, `['status', id, key, 1|0]`,
   `['fx', kind, x, y, extra]` (`hitCap` `{ id, n }`: a leader's hit cancelled by 限伤 — the renderer draws nothing;
   `extra.form` = the unit's model form from then on — an enemy's `content/enemies/helpers.js setForm`, a 傀儡师's 替身 — `shared/protocol.js fxForm`),
   `['layer', playerId, bondId, n]` (n = the layers actually added, capped at 999), `['bounty', playerId, coins]`.