0.2.5-dev: community round 30 and follow-ups

- Stock: fixed trait gifts (缪尔赛思 / 变形者集群 变形同构体, 耶拉 不融冰, 蜜蜡 / 卡涅利安 迅捷作战粮, 诗怀雅 盟约之币) still occupy stock but are granted past the cap; an operator card already shown in the shop or a reward can be taken after the pool runs out and overdraws (GitHub #498, the owner's decision); item buys and item rewards keep the stock check. 洛洛 / 焰尾 sub-pools use the measured weights (PR #497 by @2321Robin; the occupancy idea of PR #496 by @2321Robin re-implemented differently).
- Boss pairs are drawn at random once per match and fixed to battle start; the Hidden Core reuses them; a leaver does not split a full pair (the owner's decision).
- 阿戈尔 in 联防 no longer re-applies the opening devour HP loss or a false revive; Touch's team medic stays for the match after its holder leaves (owner-approved, official lifetime unverified).
- Auto-skill ranges: 司霆惊蛰 S2, 斩业星熊 S3 against air, 予愿安洁莉娜 S3; 自在's post-rebirth invulnerability is also untargetable (#510); Horn / Clementia / catapult / 忍冬 first hit (PR #502 by @Xinglan233); 不融冰 / tear gas trigger on real damage.
- Bond layer-up sound in round start / draft / prep / battle (#482): one channel, a new cue replaces the old, no extra cooldown; a prep update that adds layers and activates a bond plays one cue.
- Client: multi-swing fix (#506), refresh single-flight (PR #507 by @robll-v1), LP tower keeps the settled value with a −N tick (PR #504 by @robll-v1), export with only 潜能 / 练度 changes (PR #503 by @Xinglan233), spectator rosters list DIY members (PR #505 by @Xinglan233), banned members beside bond tooltips (PR #494 by @YuriRestia), recovery action above the away overlay (PR #511 by @RainoutOvO), the /dev/sim-check.html consistency page (PR #492 by @Cloudnyco), Korean (PR #490 by @D-Roote) and English Victoria wording (PR #499 by @YuriRestia).
- golden: 45 of 287 scenarios move against 0.2.4 (roster 17, bonds 3, fields 6, matches 17, standins 2), each attributed in docs/history/0.2.5.md §30.9 / §30.16.

docs/SIM.md

@@ -164,14 +164,18 @@ the old tile-key order), so a rotated layout plays the same (test/sim/facing-inv
 directions on an open field). 余 S3's fire wall runs through his tile perpendicular to his direction (his column facing
 RIGHT / LEFT, his row facing UP / DOWN; fx `firewall.axis` = `'col'|'row'`).
 `carryState: { hpPct, sp }` restores unite helpers' operators (HP ratio, and the SP = the official 技力, stored charges
-included: `unitsEnd` reports `skill.spTotal`, reset rebuilds the charges from it, and `skill.setSpTotal` sets it again after
-the `deploy` hook — PRTS "部署完成后…技力修改至与上一阶段结束时相同", so 独行 / 黄沙罗盘's deploy-time SP does not add to it;
+included: `unitsEnd` reports `skill.spTotal`; after all initial deployments and `battleStart` hooks, HP is restored
+against the new maximum and `skill.setSpTotal` restores the charges — PRTS "部署完成后…技力修改至与上一阶段结束时相同",
+so 独行 / 黄沙罗盘's opening SP does not add to it. Positive ratios are not raised to 1% or 1 HP;
 later redeploys keep those gifts); `carryState: { sp }` a board summon
 piece's SP only (PRTS "召唤物仅修改技力"). Nothing else is carried: a skill running at the end of the own combat enters
 switched off with the SP it had left — spent at its activation, 0 for a one-charge skill (`unitsEnd.skillActive` is
 reported, never carried; community report #34 — it used to restart for free).
 Timed deployment skills (`activateOnDeploy`) trigger once from the new deployment, consuming their deployment charge;
 they do not resume the previous skill state.
+The approved Egir carry policy rebuilds opening devour attributes while skipping its repeated HP loss on a living
+carry target; this prevents artificial knock-outs and revival-slot consumption before restoration. A fresh target
+without a previous state still takes that opening damage. Later combat, HP loss and redeploys follow the normal rules.
 `carryState: { down: true }` = an operator knocked out at the end of the helper's own combat (PRTS 卫戍协议/帮助 "上一阶段为
 退场状态的干员强制退场", user playtest #5 item 2): `start()` deploys it with everyone (initial `deploy` fires), then — before
 `battleStart` — withdraws it with reason `FORCED_EXIT` (constants.js `'forcedExit'`) and HP 0 (the end-of-phase HP ratio, as after
@@ -1065,6 +1069,11 @@ has the generic spec. Summons: `battle.tokenDef(tokenId, unit)` / `battle.spawnT
 (token skill of that slot: `variants[owner].bySkill[i]`, module attributes: `.byModule[id]`). Coverage of the selectable
 skills: `node tools/kit-coverage.mjs --missing [--tier N] [--strict]`.
 
+忍冬's 追凶 records the per-enemy start time and applies its arts addition in the same `damaged` callback,
+including the first instance (PRTS talent remarks; DESIGN §30.3). It does not refresh the ten-second window on
+repeated hits. Its `vulpisHunt` / `addition` tags and re-entrancy guard prevent the rider from feeding itself.
+`test/content/vulpis-first-hit.test.js` checks actual HP loss and damage events with naturally ready skills.
+
 Timed deployment kits: 宴 S2 uses `bb.duration` and skill ATK mods, loses current HP once on landing, and converts normal
 physical attacks to arts only while S2 is active; S1 has no such conversion. 斯卡蒂 S2 uses its own record's
 `s.bb.duration` / `s.bb.atk`; DRE-Y revival in place does not restart it. 野鬃 S1 uses its record's `duration` /
@@ -1486,6 +1495,11 @@ objects), and call `checkInvariants` at the end. Run: `node --test test/sim/*.te
 
 ## 11. Tools
 
+Equipment/fortress corrections (§30.3): Kjerag Ice and Tear Gas use actual enemy damage events, including DoT,
+not attack declarations or element-gauge/HP-loss events. Fortress fallback targeting and skill automation include
+the operator's own tile after blocked/ranged candidates; only explicitly enabled skills retain melee splash.
+`performAttack` captures each shot's `fortressMelee` flag for impact resolution, independent of later target changes.
+
 `node tools/simrun.mjs --mode mode_multi_normal --round 5 --stage act2autochess_m01 --lineup "角峰@9,7 隐现@10,4 艾丝黛尔*@12,5" --ascii 10`
 prints a per-unit damage/DPS/kills/heal table (+ ASCII frames: letters = operators, digits = ground enemies per tile,
 `^` = flyers, `■` = crates). `--wave <template>`, `--seed`, `--content generic`, `--hpMul/--atkMul/--speedMul`,
@@ -1496,6 +1510,30 @@ skill × module, every bond at its threshold and at 999 layers, every leader fie
 reduced to digests in `test/golden/*.json`; `test/golden.test.js` fails, naming the scenario and the field, when a
 digest moves. A refactor must leave them unchanged; `npm run golden:update` records an intended gameplay change.
 
+Cross-engine check (`tools/sim-check.mjs`, `public/dev/sim-check.html`): a browser simulates its battles and the server
+re-simulates them (SP_VERIFY, takeovers), so every engine must give the bits Node gives. `node tools/sim-check.mjs`
+builds the golden battle corpus (roster, bonds, fields, standins, diy — 267 BattleSpecs) and traces it in Node into
+`public/dev/simcheck/data/` (git-ignored, about a minute): after every tick the state (battle time, RNG state and per
+unit alive / removed / deployed, x, y, hp, atkCd, SP, the five element gauges, damage dealt / taken, healing, blocking
+count) hashed to the exact IEEE-754 bits, one hash per 30 ticks; every 30 ticks `b.snapshot()` (what viewers are sent);
+every 150 ticks a bounded traversal of unit-reachable state (depth 8; definitions, profile/kit, battle pointers, functions,
+accessors and `seq` left out); and `resultDigest(b.result())`. `/dev/sim-check.html` — on any device that reaches the server, a phone
+included — runs the corpus (or its fast subset, or a family) with the client's own loader (`loadBrowserSim`) and
+compares block by block: a battle that parts is reported with the first interval where a difference was detected, the level (the outcome, the
+snapshot, the state, or the hidden state only), and a download of that block's state vectors;
+`node tools/sim-check.mjs --diff <file>` replays the same ticks in Node and names differing recorded vector entries
+(with both values' bits). It cannot locate view-only/deep-only differences absent from those vectors.
+Special float values, including negative zero, survive JSON export. `--engines chromium,firefox,webkit [--subset fast|all|<family>]` runs the page headless after
+the trace (Firefox and WebKit through the `playwright` package when it is installed, `npx playwright install firefox
+webkit`; Chrome through puppeteer-core otherwise) and exits 1 on a difference. The tracing, comparison and page
+(`public/dev/simcheck/core.js`, `ui.js`, `node.mjs`) have no dependency on this repository; `stronghold.js` is the part
+that knows the sim. `--engines chrome` uses the local executable selected by `CHROME_PATH` (including Edge).
+The reference, corpus and dump share a content identity over scenarios, served sim/shared code and data. Dirty changes
+invalidate the reference. The page rechecks inputs before runs/dumps; `--diff` rejects stale or mismatched inputs.
+An empty/unknown selection, max-tick truncation or runtime exception is an error, never a zero-battle success.
+All-green means the recorded fields, sampled snapshots and result hashes agree: it does not prove equality of every
+internal field, event, projectile queue or rendered frame. Verification results belong to the tested revision.
+
 ## 12. Data notes (simdata.js)
 
 Defs are normalised from data/*.json (DATA.md) with research JSON as a fallback. Chess: `stats {maxHp, atk, def, res, aspd,