ark-stronghold-protocol/docs/history/0.2.3.md

# DESIGN §28 — 0.2.3

Part of [DESIGN.md](../DESIGN.md) (the index; section numbers are global).

## 28. New content and community fixes

This chapter describes the development version, not a published release. Numerical content is rebuilt from
CN client 2.7.81, ArknightsGameData commit `17759cc8cdf37388414145d16262a3a06f3ab25e` (2026-10-08),
with mechanism notes from PRTS checked on 2026-10-09. Launch-day descriptions may receive corrections.

### 28.1 克莱门莎 — `sim/content/kits/ops/op-clemnt.js`

- **Source:** [PRTS 克莱门莎](https://prts.wiki/w/克莱门莎) and the official character, skill and range tables.
  The non-collaboration 6★ roster gains Clementia (`char_4231_clemnt`), a 本源近卫: 72 DIY operators,
  89 backup units including stand-ins, 259 selectable forms. All three skills are selectable in the existing
  tier-5 / tier-6 DIY slots at their existing skill-rank rules; potential and cultivation follow the player's settings.
  She has no launch module. Her bonds follow the existing faction mapping, including 阿戈尔.
- **Talents:** each physical hit independently rolls 生死定夺's pre-DEF ATK multiplier; an enemy with current
  DEF below its initial DEF uses the higher probability. 习得性防御 reduces physical / Arts damage from the
  forward half-plane (including its boundary) or an enemy she blocks; 海怪 uses its separate reduction.
  Both talents take their potential-sensitive values from the generated data.
- **S1:** the charged attack applies its ATK scale and inflicts 侵蚀损伤 proportional to actual HP damage.
  **S2:** the cabin travels forwards at 2 tiles/s, up to 3 tiles, carrying up to 15 ground enemies with a total
  weight of 8. Its collision square is 1.1 × 1.1. Impassable terrain stops it; self-bound / teleport-immune
  units cannot ride. Its actual stopping position becomes a radius-1.5 vortex that can affect air, slows by 70%, and each
  second deals 200 erosion damage plus the skill's physical damage. Erosion follows the engine's normal
  resistance, burst cooldown and mode-leader immunity rules.
  Ordinary skill attacks hit their main ground target and up to four neighbours.
- **S3:** attacks select up to three ground enemies and add erosion based on actual HP damage without spending
  ammunition. A selectable ground enemy's erosion burst anywhere in the current battle marks it for 1 s.
  The closest mark consumes one of 10 bombardments: three physical + elemental hits in the target tile's
  five-tile cross. A tile with an outstanding projectile cannot be selected again for 5 s. Values are taken
  from the skill rank, not inferred from the final rank's description.
- **[ASSUMED], pending PRTS confirmation:** S2 ordinary splash radius 1.5 and stable spawn-order tie breaks;
  放逐 presently prevents walking, attacking and blocking while keeping the passenger damageable; cabin
  movement updates every sim tick, stops 0.4 tiles before an impassable edge, and the vortex ends with S2.
  S3's first impact is 0.5 s after launch, the others 1 s apart; launched projectiles finish after skill end.
  These are local kit assumptions, not changes to displacement or targeting rules for other operators.
- **Tests:** `test/content/op_clemnt.test.js` covers available forms, potential, crit probabilities, directional
  defence, S1 cadence, S2 capacity / immunity / terrain / air vortex / target cap, and S3 ammo / global marks /
  mixed damage. Golden changes and the final integration verification are recorded below once complete.

Clementia S3 currently uses the shared blast visual as an explicit `[ASSUMED]` placeholder;
this preserves the previous fallback appearance, not a claim to reproduce her official effect assets.

### 28.2 New modules — 黍「好时节」and 乌尔比安「前路」

- **Sources:** [PRTS 黍](https://prts.wiki/w/黍), [PRTS 乌尔比安](https://prts.wiki/w/乌尔比安), official
  `uniequip_table` and `battle_equip_table` from the same CN revision. Regenerating data adds GUA-Y
  (`uniequip_003_shu`) and CRU-Y (`uniequip_003_ulpia`) to their existing forms, including the anchor's
  module lookup. The day's new 5★ operators stay outside the established non-collaboration 6★ DIY pool.
- **黍:** GUA-Y's trait gives 15% damage reduction. The upgraded 天有四时 takes the improved HP / ASPD /
  four-Sui squad bonuses from the module blackboard. Satisfying at least two of the three conditions grants
  the separate additional ATK bonus (3% / 5% at module levels 2 / 3); it expires when the conditions stop holding.
- **乌尔比安:** CRU-Y increases damage to an enemy blocked by any ally by 10%. Its upgraded talent also counts another
  深海猎人's credited kills (including delayed hits after that hunter leaves) while 乌尔比安 is deployed, shares the module's specified HP / ATK proportion,
  and adds one block at 9 stacks. Potential's extra maximum stack is retained. The mode's tier-5 elite form
  still uses module level 1; support for upgraded talent blackboards does not change that established rule.
- **Tests:** `test/content/new-modules-023.test.js`, plus existing 黍 and tier-5 kit cases. The level-3 talent
  fixture is sourced from the official module blackboard; it is not exposed as a new tier-5 loadout.

### Content verification checkpoint

The new content's targeted tests passed 16/16; existing Shu / tier-5 coverage passed 120/120. An independent
interaction pass checked retreat cleanup, sleeping passengers, a real rebirth enemy, hit-count life,
post-retreat bombardments and delayed hunter kill credit. It found and corrected vortex-position rounding,
an unsupported elemental-resistance bypass and an unnecessary on-field check on the credited hunter.
All 285 golden scenarios were regenerated: only `roster` and `diy` changed, (`roster-019`, `roster-038`, `diy-059`–`diy-140`), because the added module expands
loadout coverage and the added DIY operator / Shu module expand and regroup the DIY scenarios. `bonds`,
`fields`, `matches` and `standins` remained byte-identical. Full integration and browser checks follow separately.

### 28.3 Clipboard export inside dialogs — #433

Reported by @RushFTK: exporting settings over an insecure/LAN connection focused the language selector and
reported success without copying the export. The fallback textarea was appended outside the active dialog,
whose focus trap redirected selection before `execCommand('copy')`.

`ui/clipboard.js` now selects its temporary textarea inside the active dialog/guide, validates focus and the
whole selection before copying, and always removes it and restores focus, including after exceptions.
The secure Clipboard API remains first choice; a failed fallback returns false for the existing manual-copy UI.
Shared invite/loadout copying uses the same helper. Device tests passed 16/16; independent interaction probes
passed 12/12. The real nested-dialog regression is included in `test/ui/modal.e2e.test.js` for integration testing.

### 28.4 蕾缪安's wanted range also triggers skills — #428

Reported by @CurlyZhu. Marked enemies' tiles extended normal targeting, but were absent from the skill's
initial trigger range: full SP could sit idle while she attacked a remote wanted enemy. Her kit now registers
those tiles through the existing `addTriggerRange` contract, retaining target selectability and operation cooldown.
[PRTS 蕾缪安](https://prts.wiki/w/蕾缪安) also clarifies that a new detection timer needs a selectable target,
a timer already attached continues while sleeping within detection range, and the source leaving clears its
timers and marks. Each source now tracks its own progress; one Lemuen leaving does not erase another's marks.
The focused 12 tests cover all three skills, cross-player detection in a joint field, sleep/wake, multiple sources,
redeployment (including immediate range cleanup before the next attack) and the existing S3 waiting/late-bombardment behavior. Independent review passed after fixing that cleanup.

Golden: all 285 scenarios recomputed. `roster-016/018/035/036/045`, `boss-boss_1/3/5/7-pair`,
`hidden-boss_9-pair` and `solo-HARD-9-boosted` move because wanted marks now drive skill casts and
source-owned detection/cleanup; allies and kill timing consequently change. Bonds, stand-ins and DIY are unchanged.

### 28.5 远牙's granted trait on substitute switches — #415 / PR #429

Ported from [PR #429](https://github.com/sganggs/Stronghold-Protocol/pull/429) by @CXUtk; reported and
original-mode interaction confirmed by @rosexxxp. The official activity table supplies the amounts:
`garrison_148_a/b` grants `garrison_108_a/b`, +4/+8 to active 卡西米尔 and 精准, capped at 24/48 per battle.
Only that granted trait additionally listens to `dollSwap`, in both directions, through the existing gain function.
Deployment and switches share the original cap, active-bond checks and grant deduplication. Other deployment
traits keep their existing triggers. The 7 regression cases cover both real dollkeepers, natural and S2 returns,
shared caps, duplicate grant holders, inactive bonds and unrelated timed windows.

### 28.6 Asset download visibility — PR #421

Ported from [PR #421](https://github.com/sganggs/Stronghold-Protocol/pull/421) by @LimitlessHPPK.
The asset resolver records every planned leaf with no alternative on disk as `droppedLeaves` in its report;
the terminal prints a bounded summary. `--strict` / `SP_ASSETS_STRICT=1` returns exit 1 for those omissions.
The default remains a warning, and the existing manifest shrink guard remains in force. Strict mode changes
the outcome code; it is not a transaction and does not roll back a manifest already written. Temporary test
assets are cleaned up after each case. This is an opt-in tool check, not an added packaging requirement.

### 28.7 Korean wording and contribution template — PRs #414 / #432

[PR #414](https://github.com/sganggs/Stronghold-Protocol/pull/414) by @D-Roote improves 292 existing Korean
interface strings. Only its message values are ported; later keys and the pack's translation metadata remain.
[PR #432](https://github.com/sganggs/Stronghold-Protocol/pull/432) by @Xinglan233 supplies a contribution
checklist covering the problem, source evidence, assumptions, actual checks and golden changes. The template
is installed at `.github/pull_request_template.md`, the default template location, so GitHub can offer it
without a named-template query parameter. No gameplay rules change.

### 28.8 Stop recalculating held Spine poses — PR #434

Ported from [PR #434](https://github.com/sganggs/Stronghold-Protocol/pull/434) by @Cloudnyco. A finished,
non-looping, unmixed death pose or frozen model can reuse its posed skeleton and cached pixels while its
clock advances. First poses, new clips, tint changes, mixing, multiple tracks, pending wind-ups / closing clips,
clipping switches, dirty textures and zoom rescaling still update. Redeployment resumes normal animation.
Redeploy countdown labels share rasterized textures keyed by text, colour and resolution; late web fonts refresh
them in place. The 64 targeted renderer tests passed. The author's performance measurements are not reproduced
here; real Pixi checks are part of the integration pass. No battle rule or golden result is intentionally changed.

Real Chrome verification additionally renders one actual Spine death against continual forced updates through
mixing and the held final frame (40 pixel comparisons over 600 frames), checks revival and the actor clock under the held-update policy,
and reloads Bender after a shared label was rasterised. The label updates in place and DPR variants stay separate.
This verifies correctness on that model; it is not a new performance benchmark or a full-roster pixel claim.

### 28.9 Preserve facing during pushes and pulls — #418 / PR #427

Ported from [PR #427](https://github.com/sganggs/Stronghold-Protocol/pull/427) by @YinqiMargatroid.
Successful pushes/pulls mark their displacement events `keepFacing` so the renderer retains pre-hit facing
while waiting for the endpoint snapshot and throughout the slide. Raw displacement and explicit
`keepFacing: false` retain the old directional behavior. Position, distance and unbalance time are unchanged.
The contributor's source clips and their limitations are retained in research 13; those observations were not
independently replayed for this port. **[ASSUMED]** other push sources and pulls against the target's old facing
share the examples' display behavior; the drag template alone cannot establish that engine rule.
Tests cover real push/pull skill paths, default/override metadata, snapshot reversal, and an earlier attack
arriving after slide prestart in either field direction. Golden metadata deliberately omits the display flag.

### 28.10 Charged guardian heals and mixed-skill activation — #406

Reported by @ChengCC0101. 塞雷娅 / 黍 S1 previously required an enemy to supply the attack that their charged heal
replaces. [PRTS 塞雷娅](https://prts.wiki/w/塞雷娅) and [黍](https://prts.wiki/w/黍) describe an eligible ally at
half HP or below, with the normal attack cadence. `updateAlly` now checks this alternate target after its cooldown
and control/disarm checks; an enemy is unnecessary. An invalid pending heal still returns its charge. This corrects
the enemy requirement inferred in historical §27.31; that earlier conclusion is superseded.

[PRTS 凯尔希·思衡托](https://prts.wiki/w/凯尔希·思衡托) S2 can attack enemies and heal allies. Its automatic
activation now accepts either target kind on its expanded range, under the existing player-approved range strategy
and operation cooldown. S1/S3 keep their existing trigger rules. This addresses the healing/activation parts of #406;
the separate fatal-hit/silence report still needs an independently justified damage-order fix.

Verification: 104 targeted tests pass. Independent interaction review found and verified fixes for a one-frame
cadence delay and premature charge use while disarmed; normal/100-ASPD intervals are exactly 1.2/0.6 s.
Sleep, silence, no-heal targets, charge refunds, cooperative healing and S2 target selectability remain covered.

Golden: all 285 scenarios recomputed; `diy-059`, `diy-133` change because
charged heals now act without enemy targets and S2 can open for enemies. Other scenarios remain byte-identical.

### 28.11 Integration checks for the expanded roster

The fixed roster assertions now include 72 owned DIY operators, 259 forms, 192 voiced operators and 2688 Japanese
lines; the official-table comparison still checks every potential, skill and module. Clementia adds one 阿戈尔 pick.
The official EN overlay has no launch-day Clementia talents yet: overall coverage remains ≥97%, with the measured
talent floor explicitly updated to 94.9%; untranslated text keeps its Chinese fallback. The old Saria no-enemy
assertion is corrected to require actual healing, and development builds retain the standard top-of-README notice.


## §28.12 浏览器验收退出时释放自有日志管道

完整浏览器验收中，真实对局与观战的断言已经全部通过，但 Chrome 的崩溃报告/更新辅助进程继承了日志管道，测试进程仍在等待读取结束。共用测试客户端现在在浏览器子进程明确退出后释放它的 stdio；关闭失败、子进程仍活着时不关闭管道，也不操作其他浏览器进程。游戏运行时代码与战斗规则不变。

新增真实子进程继承 stderr 的回归，以及关闭失败/信号退出的边界检查；验收要求原异常文件单独重跑并确认正常结束。首轮异常与复测结果分别保留在本轮验收记录中。

同一轮的版本重连测试在刻意停服时，Chrome 可能报告连接拒绝、连接重置或未连接的 socket。测试只在主动停服至恢复握手的窗口内排除这些已知传输日志；正常运行期间仍检查它们，脚本异常始终保留，所有版本指纹、重连及换局刷新断言不变。

### 28.13 Special targets respect DEFAULT attack timing — #438–440

Seen: 死芒's normal attack reaches enemies blocked by his own summons, but S1/S2 did not open for those targets.
The earlier 蕾缪安 fix also registered a per-tick trigger, allowing an early cast before her attack countdown elapsed.
The [PRTS mode help](https://prts.wiki/w/卫戍协议/帮助) describes DEFAULT activation at the next ordinary attack.
`addTriggerRange` now accepts `attackOnly`; both operators use it, while existing summon-area triggers retain their
per-tick behavior. Only live owned summons / wanted marks contribute selectable targets. S2's actual link range is
unchanged. PR #438 by @YinqiMargatroid addresses the same wanted-target symptom; the existing registration API is
extended instead of introducing a second targeting flag. SIM.md and design/engine.md contain the current contract.

Tests: real summoned-bone blocking, S1/S2 and united fields, attack countdown, sleep/ownership, wanted-target S1–S3,
plus existing 令 / 麦哲伦 trigger regressions. The combined operator checks pass 27 tests; the additional area-trigger
and champagne controls also pass. Golden movement is recorded below.

### 28.14 Interface text size — #435 / PR #446

Adapted from @7eLaryn: four persisted text sizes use `--t` without changing the rem layout root. All shipped language
packs retain their current translations. Added the missing explanatory note and fixed English XL settings overflow:
segmented controls and header actions wrap within their available width. Long briefing alliance labels size and wrap
with their cards instead of overlapping neighboring names. Default sm preserves the previous sizes.

Tests: settings / text-unit checks, phone landscape growth and unchanged board/canvas/shop geometry, reload persistence,
and XL settings in Chinese, English, Japanese, Korean and Traditional Chinese. The first multilingual check exposed
English overflow; the corrected browser suite passes. Current behavior is in design/client.md and PLAYING.md.

### 28.15 Cooperative card purchasing — PR #407

Adapted from @IceCodeNew. [ASSUMED] A tier V/VI shared-pool card matching another living teammate's active opening
bond receives one 18-point buying-score penalty. It is a soft preference, not reserved stock; normal/elite and DIY
stock semantics remain unchanged. Humans and bots receive the same cooperation. This does not add persona presets.

Tests: 14 targeted bot tests; 40-seed solo:HARD and coop4:HARD before/after benchmark is part of integration evidence.
The comparison includes the special-target timing changes of §28.13, so it must not be interpreted as an isolated
measurement of the purchasing heuristic. Current rule: design/match.md §6.6 and its follow-up contract.

### 28.16 AI teammate emotes — PR #444

Event wiring from @DDDarkstar, with corrected dispatch semantics. [ASSUMED] Cosmetic reactions are enabled only in
mixed human/AI matches, under the existing emote whitelist and cooldown; `SP_BOT_EMOTES=0` disables all hooks.
Responses skip cooling bots and stop after one reply. MVP uses this round's kills and damage, with stable order for
ties; settlement reacts once per round. Only the third chess merge reacts, then resets for the next round. A gift uses
its source playerId rather than a display-name heuristic. No RNG, gameplay decisions or timers are introduced.

Tests cover whitelist/cooldown, no human impersonation or recursive AI reply, pure-AI silence, round statistics,
settlement/merge limits, gift identities, the disable switch and actual emote/acquisition hooks. README, META.md and
design/match.md document the switch and boundary.

### 28.17 Active skill attack and impact sounds — PR #410

Adapted from @Convey123's updated PR. The current official `audio_data.json` gives 司霆惊蛰 mode attack banks and
赤刃明霄陈 S3 separate swing/impact banks. The generator adds per-slot `attacks` / `hits`, each with its own optional
`attackMix` / `hitMix`. It preserves the selected bank's chance and volume rather than applying the normal attack's
mix. A late impact keeps the mode at launch; an end before spawn cancels a pending start. Snapshot restoration and
ordinary-mode fallback remain supported. JP voices and current roster assets survive regeneration.

[ASSUMED] Where no activation sound names the mode, slot 1/2/3 is inferred as d/h/s. Bank selection prefers START to
ON and then plain/numeric order; this approximates an ability graph and is not independently established for every
operator. Mixed-letter banks are rejected. Instant effects that emit no normal attack event remain outside this
mechanism. ASSETS.md and design/client.md specify the manifest/runtime contract.

Tests: 85 audio, manifest, official-mix and JP-voice checks, including independent mode gain, delayed impacts,
end-before-spawn, late join and field reset. Regeneration is additive; no existing manifest leaf was removed. Four
unresolved upstream candidate icon/avatar leaves still make strict asset fetching report failure; they are not new
sound failures or deletions of existing entries. Packaging must retain this limitation visibly.

### 28.18 Per-operator voice language — #436

Roster details and selected DIY cards offer 中文 / 日本語 / 跟随全局. Settings persist a sanitized charId map locally;
normal/elite/DIY forms of that charId share the choice. Existing profiles migrate to an empty map, so global language
continues to apply. Deleting an override restores the current global choice. Missing Japanese slots / files keep the
existing Chinese fallback. No room protocol or exported combat-loadout data changes.

Tests cover malformed/inherited entries, persistence, changing the global default, deleting an override and fallback;
a real phone-sized browser changes the control, reloads and resets it. Current contract: design/client.md / PLAYING.md.

### 28.19 Browser installation — PR #413

Based on @Cromemadnd's proposal, with a one-use install controller. Buttons appear only after a browser install offer;
standalone/fullscreen and installed sessions hide them. Accept, dismiss, failure and duplicate clicks all consume the
same offer once. See [the prompt API](https://developer.mozilla.org/en-US/docs/Web/API/BeforeInstallPromptEvent/prompt)
and [installability requirements](https://developer.mozilla.org/en-US/docs/Web/Progressive_web_apps/Guides/Making_PWAs_installable).
The existing original fortress favicon supplies correctly sized 192/512 px icons, shipped outside optional game art.
No service worker or offline cache is installed; the server is still required.

Tests: offer lifecycle/disposal, actual PNG dimensions, real Chrome manifest and installability checks, and a synthetic
dismissed offer through the actual button. Initial incognito rejection was a test-environment limitation; the real
installability check runs in Chrome's ordinary temporary profile. Native OS installation itself is not performed.

### 28.20 Recover a closed local match window — #431

Reproduced with an idle window retaining its own session while a second match window closes. The title/lobby now
lists local saved match seats using bounded metadata beside the existing recent-token ring. Explicit selection asks
live tabs to claim the chosen token again before adopting it and reloading. An active tab is never deliberately
replaced; names/IP are not authentication. Missing BroadcastChannel refuses recovery. A server-generated replacement
token or leaving clears obsolete metadata. No server API changes; normal automatic startup selection is preserved.

Tests: live tab protection, closed-tab recovery, simultaneous claim arbitration, no channel/unknown id, malformed
storage and expired metadata; a real two-window browser restores the identical server playerId and room after closing
the match tab. The existing identity/static suite also passes. Current contract: design/client.md and network.md.

### 28.21 Unanimous opening reroll — #77 / PR #408

Adapted from @Siornya, excluding the unrelated result-screen redesign. [ASSUMED] This remake feature is host-proposed,
requires every connected human's consent, and pauses confirmation while voting. Bots/spectators do not vote; autoplay
humans still do. Single-human matches refresh immediately. Refusal, cancellation, disconnect or leave restores the
old setup and remaining confirmation time. Success preserves room/seats/settings, resets Ready and restarts the full
normal timer; fixed stages/bans remain fixed. The random result may repeat.

A dedicated seed-derived reroll stream does not advance the original shop/draft/meta streams. A 300 ms anti-burst guard,
setup revision and vote id reject stale actions. Opening data and every player's DIY stock are prepared before any
mutation; a forced failure on the second player's stock leaves all previous state intact. Current contracts are in
META.md, design/match.md, design/network.md and PLAYING.md; all four translation packs have the new messages.

Tests: 13 match checks including real four-player sockets, permission/spectator checks, cancellation/disconnect,
reconnect, last-second and queued-Ready timers, outdated messages, one human with bots, fixed rules, fuzzed requests
and transactional failure. Final browser/integration validation supplements these targeted checks.

### 28.22 Late-loading deployment animation — part of PR #416

@Cromemadnd's proposal exposed a real missing case: onDeploy arrived before the model, so Start was never played.
The view now retains elapsed game time until the model arrives and continues the existing clip at that point. A clip
already over is not restarted; an intervening attack/death cancels the pending entrance. Existing Front/Back transfer
and single-source deployment audio remain unchanged. The proposed repeated phase/purchase sound sequence is not part
of this fix: its timing and duplication need separate visual evidence.

Tests: delayed load, simulation rate, load after clip end, attack/death supersession, destroyed view and existing
UnitView behavior, 48 checks passing. No simulation/golden change from this visual fix.

### 28.23 Follow-up golden movement

Recomputed all 285 scenarios from the integrated code; never copied PR baselines. The DEFAULT cadence correction moves
`roster-016`, `roster-035`, `roster-036`, `roster-045`; paired fields `boss-boss_1-pair`, `boss-boss_3-pair`,
`boss-boss_5-pair`, `boss-boss_7-pair`, `hidden-boss_9-pair`; and match `solo-HARD-9-boosted`. 死芒's added activation
condition moves `diy-132`. `coop2-NORMAL-14-standins` includes both cooperative purchasing and the corrected simulation.
The pulse queue correction of §28.24 additionally moves `hidden-boss_10-solo`. The bonds and standins families are unchanged. Cosmetic emotes, preferences, installation and opening rerolls that
are not requested do not intentionally alter those games. The validation record lists any outcome flips and measured
benchmark changes; the independent final review remains pending.

### 28.24 Echo / counter recursion found by the cooperative benchmark

The 40-seed benchmark exposed existing hook-depth errors in co-op seeds 3 and 29, plus seed 14 after the purchasing
change. In all three, “余音” damage pulses and 刺玫 / 泡泡 counters synchronously called each other. A scaled echo can
have 67 hit-count HP; the generic 32-hook guard cut off effects before the legitimate chain finished. The same chain
could also enter another hit while zero HP was waiting for death bookkeeping.

[PRTS “余音”](https://prts.wiki/w/“余音”) establishes the hit-count and damage-pulse behavior. [ASSUMED] The exact nested
ordering is not documented: each echo now queues recursively earned pulses and resolves them FIFO in the same call,
after the current pulse finishes. Every earned pulse retains its attack value, including the lethal hit's pulse. No
extra tick, RNG draw, global hook-depth increase or arbitrary pulse suppression is introduced. Death prevents further
counter targeting normally. This local scheduling choice needs final fidelity review.

Tests use real elite 刺玫 and 泡泡 against 67-hit echoes: all 67 pulses resolve, the echo dies, no engine error is logged,
and the queue clears. Both tests fail against the previous implementation. Together with the existing enemy/boss and
刺玫 healing suites, 273 tests pass. The benchmark now writes actual diagnostic messages beside nonzero error counts,
so failures cannot be mistaken for ordinary performance results. Recomputed golden and final benchmark results are
recorded with the candidate; do not reuse the earlier, erroneous benchmark as a clean result.