0.2.6-dev: the WebSocket compression figures in the code and docs are the server counters' measurement (173 KB of frames → 60 KB on the wire)

docs/DEPLOY.md

@@ -9,7 +9,7 @@
 |---|---|
 | 服务器 CPU | 战斗在各玩家浏览器里模拟（DESIGN §14），服务器只负责回合、经济和校验：**每个房间每个作战回合约 1 ms CPU**。AI 队友 / 掉线玩家的战场由服务器模拟：作战开始时 3 个 AI 战场在开发机上约 0.2–0.5 s CPU，小主机上可能要几秒（分成 8 ms 小片执行，不会卡住其他房间）。`SP_VERIFY=all` 会复算每个真人战场，CPU 明显增加，小主机建议保持 `off` 或 `sample`。 |
 | 服务器内存 | 空闲约 100 MB，每个进行中的对局再增加几 MB。 |
-| 网络 | 4 人对局中服务器每回合下行约 0.25 MB（DESIGN §14 实测；0.2.6 起 1 KB 以上的 WebSocket 消息用 permessage-deflate 压缩，实测一局单人对局的服务器消息从 169 KB 降到 63 KB，约为原来的四成）。首次进入游戏时浏览器要从主机下载所需的图片 / Spine 模型 / 音频（按需加载，之后走浏览器缓存），公网隧道带宽小时第一次会慢一些。0.2.6 起服务器启动后在后台给素材算内容哈希（第一次约 1–20 秒，看机器和硬盘；之后只检查大小和修改时间），下发的素材清单里每个地址带上哈希：浏览器把图片、模型和音频缓存一年，服务器更新、重新解压或 Docker 重建之后，玩家只重新下载内容真的变了的文件。哈希存在 `.cache/asset-hashes.json`，删掉会重新计算；目录不可写（例如只读的容器）时只在内存里算。 |
+| 网络 | 4 人对局中服务器每回合下行约 0.25 MB（DESIGN §14 实测；0.2.6 起 1 KB 以上的 WebSocket 消息用 permessage-deflate 压缩，实测一局单人对局的服务器消息从 173 KB 降到 60 KB，约为原来的三分之一）。首次进入游戏时浏览器要从主机下载所需的图片 / Spine 模型 / 音频（按需加载，之后走浏览器缓存），公网隧道带宽小时第一次会慢一些。0.2.6 起服务器启动后在后台给素材算内容哈希（第一次约 1–20 秒，看机器和硬盘；之后只检查大小和修改时间），下发的素材清单里每个地址带上哈希：浏览器把图片、模型和音频缓存一年，服务器更新、重新解压或 Docker 重建之后，玩家只重新下载内容真的变了的文件。哈希存在 `.cache/asset-hashes.json`，删掉会重新计算；目录不可写（例如只读的容器）时只在内存里算。 |
 | 磁盘 | 素材约 550 MB（`public/assets`，含中文约 65 MB、日文约 85 MB 两套干员语音）+ 依赖约 125 MB（`node_modules`；整合包只带运行依赖，约 65 MB）；可选的本地提取约 40 MB（`.venv-extract`）+ 70 MB 贴图（见第 6 节）。完整包解压后约 710 MB。 |
 | 玩家设备 | 支持 WebGL 的现代浏览器（Chrome / Edge / Firefox / Safari 最新版），电脑或手机平板（横屏）。老旧设备可在设置里调低画质或访问 `/?board=2d`。 |
 

docs/design/network.md

@@ -4,7 +4,7 @@ Part of [DESIGN.md](../DESIGN.md) (the index; section numbers are global).
 
 ## 8. Network protocol (shared/protocol.js is normative)
 
-Transport: one WebSocket per tab at `/ws`, JSON text frames `{ t: '<type>', ...payload }`. Client requests may carry `rid` (request id); the server's direct reply echoes `rid` (`ok` or `error`). Server pushes have no `rid`. Max frame 64 KB inbound (a compressed frame counts its inflated size); rate limit 40 msgs/s per connection (excess dropped with `error RATE`). Frames from 1 KB up are compressed with permessage-deflate (RFC 7692; `server/http/websocket.js WS_DEFLATE`: no server context takeover, a 2 KB server window, zlib memLevel 7 — about 72 KB of zlib state per connection); smaller frames go out plain, and a client that offers no compression, or caps the server window below 2 KB, gets plain frames. A measured solo session (rounds 3–5): 169 KB of server frames → 63 KB on the wire.
+Transport: one WebSocket per tab at `/ws`, JSON text frames `{ t: '<type>', ...payload }`. Client requests may carry `rid` (request id); the server's direct reply echoes `rid` (`ok` or `error`). Server pushes have no `rid`. Max frame 64 KB inbound (a compressed frame counts its inflated size); rate limit 40 msgs/s per connection (excess dropped with `error RATE`). Frames from 1 KB up are compressed with permessage-deflate (RFC 7692; `server/http/websocket.js WS_DEFLATE`: no server context takeover, a 2 KB server window, zlib memLevel 7 — about 72 KB of zlib state per connection); smaller frames go out plain, and a client that offers no compression, or caps the server window below 2 KB, gets plain frames. A measured solo session (rounds 3–5, `/healthz?traffic=1`): 173 KB of server frames → 60 KB on the wire.
 
 ### 8.1 Session & lobby
 C→S: `hello {name, token?, version}` · `room.create {mode:'solo'|'coop', difficulty}` · `room.join {code}` · `room.leave` · `room.ready {ready}` · `room.setDifficulty {difficulty}` (host) · `room.setAiPicksLast {on}` (host, co-op, before the match — `ROOM_STARTED` during it, `BAD_TARGET` in a solo room; un-readies the other humans like a difficulty change; the match takes it as `opts.aiPicksLast`: humans draft before AI seats, off by default, no extra random draw) · `room.addBot` / `room.removeBot {seat}` (host) · `room.kick {seat, playerId}` (host, before the match: another human leaves the room and gets `room.closed {reason:'kicked'}`; may join again; `playerId` = the player the host confirmed, a seat that changed hands meanwhile → `BAD_TARGET`) · `room.start` (host) · `room.loadout {entries, ops?}` (§16: any time in the lobby / room and during INFO_CHECK, then `WRONG_PHASE`; heavy rate-limit bucket; the match exposes it as `m.private.loadout`; `ops` — 0.2.2 — the per-operator 潜能 / 练度 `{ [charId]: { potential?: 1–6, cultivate?: 0–3 } }` of the 干员调配 roster and the 自选 owned pool, ≤ 256, checked strictly with the entries, `m.private.ops`; absent = none set: every operator 潜能 6, 精英2 Lv.60) · `room.ownership {notOwned}` (0.2.0 补位, the 干员持有 tab: the base chess ids marked as not owned — ≤ 160 ids, kept leniently: anything but a droppable NORMAL chess is dropped; stored on the session and the seat, a match takes the seat's list when it starts and never changes it — during a match `ROOM_STARTED` 'stored for the next match'; heavy bucket; the match exposes it as `m.private.standIns`) · `room.diy {picks}` (0.2.0 自选编队: `{ [slotBaseId]: { charId, skillIndex?, uniEquipId? } | null }` for the four DIY slots — ≤ 8 entries, kept leniently: shared/protocol.js `checkDiyPicks` drops an illegal pick (an operator without a kit, another tier's prototype, a prototype off its locked skill, a module of another game mode (集成战略 / 生息演算), an owned operator's second slot, the same operator twice in a tier, an unknown slot), only malformed picks are `BAD_MSG`; stored on the session and the seat like the not-owned list, a match takes the seat's picks when it starts — during a match `ROOM_STARTED` 'stored for the next match'; heavy bucket; the match exposes them as `m.private.diy`) · `ping {c}`

docs/history/0.2.5.md

@@ -480,9 +480,10 @@ Layer 0:
   and voices were already gated. Normative line: design/client.md Audio.
 - **WebSocket permessage-deflate** (`server/http/websocket.js WS_DEFLATE`): frames from 1 KB, no server context
   takeover, a 2 KB server window, zlib memLevel 7 (about 72 KB of zlib state per connection). [ASSUMED] the parameters
-  come from one measured solo session (216 server frames: 169 KB → 63 KB; windows above 2 KB gained nothing); co-op
-  messages are the same kinds. A client that offers a narrower server window than 2 KB, which ws would refuse with a 400,
-  connects uncompressed instead (`narrowDeflateOffer`). Normative line: design/network.md §8 Transport.
+  come from measured solo sessions (the server counters: 173 KB of frames → 60 KB on the wire; replaying 216 captured
+  frames, windows above 2 KB gained nothing); co-op messages are the same kinds. A client that offers a narrower server
+  window than 2 KB, which ws would refuse with a 400, connects uncompressed instead (`narrowDeflateOffer`). Normative
+  line: design/network.md §8 Transport.
 - **玩法说明 pages as WebP.** `tools/local-extract/extract.py` writes a WebP copy beside each of the 19 local pages and
   the manifest lists it (`GUIDE_WEBP_MODE`); the guide preloads only the next page instead of both neighbours. The
   copies are lossless (25.6 MB → 17.4 MB on the CN client's pages). Lossy quality 90 would make them 4.2 MB; per the

server/http/websocket.js

@@ -19,9 +19,10 @@ export const WS_MAX_PAYLOAD = 64 * 1024;
  * permessage-deflate (RFC 7692) for the JSON text frames. A message from 1 KB up (m.public, m.private, b.start, welcome
  * …) is compressed on its own — no context is kept from one message to the next (`serverNoContextTakeover`) — with a
  * 2 KB window and zlib memLevel 7, so a connection holds about 72 KB of zlib state; smaller frames (pong, ok) go out
- * as they are. Measured on a solo session (rounds 3–5, 216 server frames): 169 KB of payload → 63 KB on the wire; a
- * larger window gains nothing on these messages. Browsers negotiate it by themselves; a client that offers none gets
- * plain frames. The inflated size of a client frame still counts against WS_MAX_PAYLOAD.
+ * as they are. Measured on a solo session (rounds 3–5, the server's own counters, traffic.js): 173 KB of payload →
+ * 60 KB on the wire; replaying a session's frames with larger windows gained nothing on these messages. Browsers
+ * negotiate it by themselves; a client that offers none gets plain frames. The inflated size of a client frame still
+ * counts against WS_MAX_PAYLOAD.
  */
 export const WS_DEFLATE = Object.freeze({
   threshold: 1024,