修复赤刃明霄陈三技能对空自动开启并核实社区反馈

S3 的扩展范围含可选中的飞行目标时可以自动开启剑气；普攻保持仅对地。
[ASSUMED] 自动触发区域沿用扩展攻击格和永久射程延长，不扩张到未来剑气的整条路线。

补测部署技能与休整属性预览、后排协防蕾缪安真实锁定耗弹、诗怀雅落地金币、
莎草对炸弹的治疗链，以及予愿三种技能的普通地面溅射和无视可选性例外。
未复现的报告保留证据边界；诗怀雅被动落地投瓶符合 PRTS。

全 285 场 golden 重建并验证，仅 diy-074 变化：新增 4 次陈 S3，伤害 142376→163563，
时序/选敌改变后击杀 21→18、漏怪 96→99，223.83 秒和结束方式不变。
其余 284 场不变。

CI：5879 tests，5876 pass，0 fail，3 skip；setup/smoke/lint/imports/typecheck 均通过。
新增 16 项回归，相关 6 文件 63/63，补充 2 文件 20/20，文档 31/31。

docs/history/0.2.3.md

@@ -395,3 +395,25 @@ Full golden regeneration covers all 285 scenarios. Four digests intentionally mo
 
 No scenario's ending reason changes. The bonds, matches and DIY families are byte-identical to the preceding golden
 files. The 50-second setting and fallback-message change do not alter battle rules in those digests.
+
+### 28.26 Further community reports: deployment, air targets and champagne
+
+Five reports were checked against the current candidate and PRTS on 2026-10-10. Reports without an operator,
+phase or enemy identity remain bounded by the scenarios actually tested.
+
+| Report | Evidence and result |
+|---|---|
+| Deployment skill counts down before combat | Not reproduced in the battle lifecycle or real match prep-stat path. Constructing a battle does not deploy units; starting it activates deployment skills. Prep previews start their own battle for initial stats, never step it, send no SP/skill timer and cannot spend the next battle's duration. Tests cover 予愿安洁莉娜 S1 and repeated prep previews with 缄默德克萨斯. This does not identify the unspecified UI/phase in the report. |
+| 赤刃明霄陈 will not open S3 against flyers | Reproduced and fixed. `ACTIVE_RANGE` checked her ground-only normal-attack profile, although the sword wave can hit air ([PRTS](https://prts.wiki/w/赤刃明霄陈)). S3 now registers air-capable targeting on its expanded attack grid. Normal swings remain ground-only; operation cooldown, stealth/sleep selectability and stun still apply. [ASSUMED] The auto-cast region follows the existing expanded-grid policy, including permanent range extension, rather than the wave's whole future route. This is an automation choice, not a claim about the official mode's automatic-cast rule. |
+| Rear player's 蕾缪安 does not open S3 on front player's targets | Already handled by §28.4/§28.13. Added normal/elite united-field cases with the second player's 蕾缪安 and the first player's distant Laterano operator: the real wanted timer completes, S3 starts automatically, locks that enemy and spends ammo. The existing all-skill tests retain attack cadence and selectability. |
+| 琳琅诗怀雅 S2 immediately creates a bottle on deployment | Expected. [PRTS](https://prts.wiki/w/琳琅诗怀雅) identifies S2 as passive, deployment with S1/S2 as skill activation, and her talent as granting a coin on activation. The first bomb spends that coin without waiting for a merchant payment. No new SP cooldown is added. Existing throw-animation timing remains an assumption. |
+| 莎草 shields a champagne bomb | Not reproduced on the current candidate, which already gives bombs 禁疗 under the maintainer's earlier decision. Both skills actually heal/shield an adjacent operator but never chain to the bomb produced by 琳琅诗怀雅. [PRTS 莎草](https://prts.wiki/w/莎草) also excludes 禁疗 from S2's main-target selection. |
+| 予愿安洁莉娜 dies to splash while airborne | All three skills avoid ordinary ground 控潮术师 splash while airborne and take it after landing. A 萨卡兹悖谬暴虐兵长 splash explicitly ignoring target selectability still damages her, as expected. 起飞's 对地规避 is a selectability effect, not immunity to every damage source ([PRTS 异常效果](https://prts.wiki/w/异常效果), [暴虐兵长](https://prts.wiki/w/萨卡兹悖谬暴虐兵长)). Without the reported enemy's identity these checks cannot settle that particular encounter. No global immunity change. |
+
+All 285 golden scenarios were regenerated and verified. Only `diy-074` moves: the mixed 陈 loadout against
+flying enemies now starts four previously idle S3 activations and emits sword-wave effects. Damage rises
+142376 → 163563; changed combat/target timing moves kills 21 → 18 and leaks 96 → 99, with the same 223.83 s
+clear time and end reason. All other 284 scenarios, including the five other golden families, are unchanged.
+The focused checks cover all four tier/elite forms, selectability/control/cooldown boundaries, actual casts,
+healing chains and enemy splash. Full CI passes: 5879 tests, 5876 passed, 0 failed, 3 skipped; setup, smoke, lint,
+imports and typecheck pass. This round changes no production UI and does not rerun the opt-in browser suites.