feat: per-player buy classification, AWP-hold rate, grenade icons & glass UI

CLAUDE.md

@@ -30,7 +30,7 @@ web_server.py  (Flask, port 5000)
     ▼
 output/player_{domain}.json  +  output/analysis_summary.json
     ▼
-Browser fetches /api/player/<domain> → ReplayPlayer canvas (9s loop, CT/T × Pistol/Full/Eco)
+Browser fetches /api/player/<domain> → ReplayPlayer canvas (looping, unified CT/T; merged-pistol + per-player Buy)
 ```
 
 ### Module Map (`server/`)
@@ -40,7 +40,7 @@ Browser fetches /api/player/<domain> → ReplayPlayer canvas (9s loop, CT/T × P
 | `maps.py` | Runtime map loader: `load_map(name)`, `game_to_pixel(transform,gx,gy)`, `available_maps()` |
 | `setup_maps.py` | One-time: pull radar.png + transform from awpy into `data/maps/<map>/` |
 | `parse.py` | `get_round_table`, `classify_rounds` (3-type CT/T), `parse_positions`, `parse_grenades_for_rounds`, `parse_demo` (merger) |
-| `combat.py` | `parse_combat_stats` (K/D + CT AWP), `aggregate_combat_stats` |
+| `combat.py` | `parse_combat_stats` (K/D + AWP-hold rate), `aggregate_combat_stats` |
 | `player_json.py` | `build(...)` — assembles per-player JSON from rounds + combat |
 | `pipeline.py` | `run(usernames, map_name, ...)`, demo dedup index, `cleanup_demos`, `download_and_extract`, `assemble_round_offset` |
 | `api_client.py` | 5E scraping: `search_player`, `get_demos_by_domain(domain, map_name, count)`, `get_steamid_for_player`, `download_demo` |
@@ -87,12 +87,18 @@ if total `.dem` size > 30 GB, deletes oldest files until ≤ 10 GB.
 - `game_to_pixel(transform, gx, gy)` = `((gx-pos_x)/scale, (pos_y-gy)/scale)` — Y axis inverted.
 - `available_maps()` lists dirs under `MAPS_DIR` that contain `meta.json`.
 
-### Round Classification (parse.classify_rounds) — 3 types, CT and T
+### Round Classification (parse.classify_rounds) — 2 kept types, CT and T
 
-At each `round_freeze_end`, for the target steamid: `side` ∈ {CT, T}. Economy:
-- **Pistol** — first round of each half-segment (side flips into a new half).
-- **Full** — team-avg `current_equip_value` ≥ `EQ_FULL_BUY` (3800).
-- **Eco** — otherwise. (No Force Buy in 2.0.)
+At each `round_freeze_end`, for the target steamid: `side` ∈ {CT, T}. Economy is judged on
+the target's **own** `current_equip_value` (not team average):
+- **Pistol** — first round of each half-segment (side flips into a new half). Always kept,
+  regardless of equip.
+- **Buy** — non-pistol with personal equip ≥ `EQ_BUY_MIN` (2000).
+- Non-pistol with personal equip < 2000 → **dropped**: `rtype=None`, excluded from JSON.
+  The round still carries its real `side` so half-segment (pistol) tracking isn't broken.
+
+Downstream `parse_positions`/`parse_grenades_for_rounds`/`parse_deaths_for_rounds` filter on
+`r["side"] and r["rtype"]`, so dropped rounds produce no path/grenades/death.
 
 ### Position Sampling (parse.parse_positions)
 
@@ -108,9 +114,11 @@ Types: smoke/flash/he/molotov/decoy. Durations: smoke 18s, molotov 7s, decoy 15s
 ### Combat Stats (combat.py)
 
 - **K/D** — global scoreboard: `kills_total`/`deaths_total` at last `round_end` tick; averaged across demos.
-- **AWP rate** — CT-side AWP kills / CT-side total kills. CT rounds via `classify_rounds` `fe_tick`/`end_tick`.
-  `SNIPER_EVENT_NAMES = {"awp","ssg08","g3sg1","scar20"}` (plain event `weapon` names).
-  Aggregated as `sum(awp_kills)/sum(ct_kills)`.
+- **AWP rate** — **AWP-hold rate**: rounds where the player ever held an AWP / total rounds played
+  (both sides, any economy). "Held" = the `inventory` tick field (weapon display names) contains
+  `"AWP"` at any sample tick across the round window. Per demo returns `{awp_rounds, total_rounds}`;
+  aggregated as `sum(awp_rounds)/sum(total_rounds)`. (Replaces the old CT-side AWP-kill ratio, which
+  was too sparse and read ~0% for most players.)
 
 ### Endpoints (web_server.py)
 
@@ -121,6 +129,7 @@ Types: smoke/flash/he/molotov/decoy. Durations: smoke 18s, molotov 7s, decoy 15s
 - `GET /api/player/<domain>` — that player's `player_{domain}.json` (404 if missing).
 - `GET /api/results` — raw `analysis_summary.json` (no path normalization in 2.0).
 - `GET /output/<file>` — serves output JSON. `GET /maps/<path>` — serves radar from `MAPS_DIR`.
+- `GET /icons/<path>` — serves grenade icon SVGs from `ICONS_DIR` (repo-root `radar/icons/`).
 - `GET /` — `index.html`.
 
 ### JSON Contract
@@ -143,7 +152,7 @@ Types: smoke/flash/he/molotov/decoy. Durations: smoke 18s, molotov 7s, decoy 15s
   ]
 }
 ```
-`side` ∈ {CT,T}; `rtype` ∈ {Pistol,Full,Eco}; all `t` are seconds from freeze_end; coords are game coords.
+`side` ∈ {CT,T}; `rtype` ∈ {Pistol,Buy}; all `t` are seconds from freeze_end; coords are game coords.
 
 `output/analysis_summary.json`:
 ```json
@@ -155,16 +164,24 @@ Types: smoke/flash/he/molotov/decoy. Durations: smoke 18s, molotov 7s, decoy 15s
 
 ### Frontend (templates/index.html + static/app.js + static/replay.js)
 
-- Sticky 50px header + left sidebar (map `<select>`, 5 username inputs, depth, key, scan button,
-  status, failed list) + right panel of player cards.
+- Sticky 50px header (with a **unified CT/T toggle** + global play/pause + scrubber) + left sidebar
+  (map `<select>`, 5 username inputs, depth, key, scan button, status, failed list) + main panel.
+- Main panel = a **merged pistol overlay** on top (`#pistol`: one canvas overlaying *all* scanned
+  players' Pistol rounds, each player a distinct color + legend) followed by per-player cards.
+- The header CT/T toggle drives `side` on **every** canvas at once (merged pistol + each card's buy
+  canvas); one side is shown at a time. There are no per-card tabs.
 - `app.js`: `loadMaps()` fills the dropdown; `run()` POSTs `/api/analyze`; `poll()` hits `/api/status`
-  every 2s and adds a card per result via `/api/player/<domain>`.
-- Each card: K/D, AWP%, round count; CT/T tabs; 3 canvases (Pistol/Full/Eco). Switching tab calls
-  `setFilter(side, rtype)` on each `ReplayPlayer`.
+  every 2s and adds a card per result via `/api/player/<domain>`. Each card: K/D, AWP-hold %, round
+  count, and a single **Buy** canvas. The merged overlay grows a shared `pistolRounds` array as
+  players load; the toggle calls `setFilter(side, fixedRtype)` on each registered `ReplayPlayer`.
 - `replay.js` `ReplayPlayer(canvas, {radar, transform, rounds, side, rtype})`:
-  overlays all matching rounds on a `PLAYBACK_S=9`s loop (45s game time, 5× speed). **No fading trails.**
-  Draws grenade in-flight arcs, landing dots, and range circles (smoke/molotov) during `[land_t, expire_t]`.
-  Methods: `setFilter`, `toggleRound`, `start`, `stop`. `static/replay_test.html` is a standalone fixture.
+  overlays all matching rounds on a `PLAYBACK_S` loop (`WINDOW_S` game time accelerated). **No fading
+  trails.** Per-round `color` overrides the side color (used by the merged pistol overlay). Draws
+  grenade in-flight arcs, landing dots, and range circles (smoke/molotov) during `[land_t, expire_t]`.
+  While a nade is airborne (`[throw_t, land_t)`) it also draws a **white SVG icon** at the arc head
+  (smoke/flash/he/molotov, served from `/icons/<file>.svg`, drop-shadow for contrast); the icon
+  disappears on detonation (no decoy icon). Source SVGs live in repo-root `radar/icons/`.
+  Methods: `setFilter`, `toggleRound`, `drawAt`, `_drawNadeIcon`. `static/replay_test.html` is a standalone fixture.
 
 ### Known Data Limitations
 
@@ -232,7 +249,7 @@ Always cast result to `pd.DataFrame` and cast `steamid` to `str`.
 
 | File | Purpose |
 |------|---------|
-| `server/config.py` | `HOST/PORT/SECRET_KEY`, paths, `MAPS_DIR`, `TICK_RATE=64`, `WINDOW_S=45`, `SAMPLE_EVERY=8`, `EQ_FULL_BUY=3800` |
+| `server/config.py` | `HOST/PORT/SECRET_KEY`, paths, `MAPS_DIR`, `TICK_RATE=64`, `WINDOW_S=20`, `SAMPLE_EVERY=8`, `EQ_BUY_MIN=2000` (per-player buy floor; `EQ_FULL_BUY` legacy/unused) |
 | `server/data/maps/<map>/radar.png` | Radar background (generated by setup_maps.py) |
 | `server/data/maps/<map>/meta.json` | Coordinate transform `{pos_x,pos_y,scale}` |
 | `server/demos_opponents/` | Downloaded .dem files + `.demo_index.json` |

docs/superpowers/specs/2026-06-25-buy-classification-awp-layout-design.md

@@ -0,0 +1,104 @@
+# 设计：个人买局分类 + 持狙占比 + 合并布局
+
+日期：2026-06-25
+状态：已批准设计，待写实现计划
+
+## 背景与目标
+
+CS-Scout 2.0 当前把回合分为 Pistol/Full/Eco（按全队平均装备值），每名对手一张卡片、
+三宫格（Pistol/Full/Eco）、每卡各自带 CT/T tabs；AWP 率口径为"仅 CT 方 AWP 击杀 / CT 方
+总击杀"，数据稀疏导致几乎总是 0%。
+
+本次改动三件事：
+
+1. 回合分类改为**个人判定**，去掉 Eco，丢弃个人装备过低的局，只保留 **Pistol** 和 **Buy**。
+2. AWP 率改为**持狙回合占比**，修掉一直 0% 的问题。
+3. 前端重构：顶部**统一 CT/T 切换**；**手枪局把 5 名对手合并到一张图**置顶；每名对手卡片只留
+   一张**买局**图。
+
+## 1. 回合分类（`parse.py` `classify_rounds`）
+
+判定点不变：每回合 freeze_end 快照，字段 `current_equip_value`。
+
+- **改个人判定**：用目标玩家个人的 `current_equip_value`，不再算全队平均。
+- 类型 3 种 → 2 种：
+  - **Pistol** — 每半场首回合（现有 `pistol_num` / 换边逻辑不变；无条件保留，不受 2000 门槛限制）。
+  - **Buy** — 非手枪局且个人装备 `>= EQ_BUY_MIN`（2000）。
+  - 非手枪局且个人装备 `< 2000` → **丢弃**：`rtype = None`，不进入 JSON。
+- **换边检测不能被丢弃局打断**：手枪局判定依赖 `prev_side`。因此即使某局被丢弃，`side` 仍按
+  真实 CT/T 计算并用于 `prev_side` 跟踪，只把 `rtype` 置 `None` 表示丢弃。否则下一局会被
+  误判为半场首局（=手枪局）。
+- **下游过滤**：`parse_positions` / `parse_grenades_for_rounds` / `parse_deaths_for_rounds`
+  当前的 `active = [r for r in classified if r["side"]]` 改为
+  `if r["side"] and r["rtype"]`，从而排除被丢弃的局。
+- `config.py` 新增 `EQ_BUY_MIN = 2000`。`EQ_FULL_BUY` 不再被 `classify_rounds` 使用（保留常量，
+  避免触碰旧引用；如确认无其他引用可一并删除）。
+
+## 2. AWP 率 → 持狙回合占比（`combat.py`）
+
+- 废弃旧指标（CT 方 AWP 击杀 / CT 方击杀）。
+- 新指标：**持狙回合数 / 总回合数 × 100**。
+- **"持狙"判定**：在该回合的采样 tick 上读目标的 `inventory`（demoparser2 tick 字段，返回武器名
+  列表），任一采样点的 inventory 含 AWP（匹配 `awp` / `weapon_awp`，大小写不敏感）即记该回合为
+  持狙回合。
+  - 实现前先用真实 demo 验证 `inventory` 字段可用及其武器名格式。
+  - **回退方案**：若 `inventory` 不可用，退化为"该回合内目标有 AWP 击杀"（player_death，
+    attacker=目标 且 weapon=awp，tick 落在回合窗口）。
+- **分母 = 该玩家出场的全部回合**（CT + T，不限经济，含被分类层丢弃的 eco 局）。即用
+  `classify_rounds` 结果里 `side` 非空的全部回合，不要求 `rtype`。
+- `parse_combat_stats` 返回 `{kd, awp_rounds, total_rounds}`。
+- `aggregate_combat_stats`：`awp_rate = round(sum(awp_rounds)/sum(total_rounds)*100, 1)`，
+  分母 0 时为 0.0。`kd` 聚合不变。
+
+## 3. 前端布局
+
+### 统一 CT/T 切换（`index.html` header + `app.js`）
+- 在 header 的 `.ctrls`（与播放/进度条同排）放一个 CT/T 切换（两个按钮，默认 CT）。
+- 该开关控制页面上**所有** canvas 的 side：手枪合并图 + 每个对手买局图。一次只看一边。
+- 移除每张卡片内的 CT/T tabs。
+
+### 手枪局合并区（`#pistol`，置于 `#cards` 上方）
+- **一张大 canvas**，叠加本次扫描到的最多 5 名对手的**手枪局**路径。
+- **每名玩家一种颜色** + 文字图例（玩家名 ↔ 颜色）。
+- 受顶部统一开关控制：显示当前 side 的手枪局。
+- 实现：构造合并轮次列表 = 各玩家 `rounds` 中 `rtype=="Pistol"` 的局，每局打上该玩家颜色 tag，
+  作为单个 `ReplayPlayer` 的 `rounds`，`rtype="Pistol"`。
+
+### 每名对手卡片（`#cards`）
+- 保留：玩家名、K/D、AWP%（持狙占比）、回合数。
+- 去掉：CT/T tabs、Pistol/Full/Eco 三宫格。
+- 只留**一张 canvas**，叠加该玩家所有 **Buy** 局，`rtype="Buy"`，按 side 上色。
+
+### `replay.js`（最小改动）
+- 支持 per-round 颜色覆盖：dot/X/arrow 颜色取 `r.color || SIDE_COLOR[this.side]`
+  （把 `col` 计算移入按 round 循环内）。单人买局图不带 `color`，仍按 side 上色；合并图每局带
+  `color`。
+- `setFilter(side, rtype)` 不变。统一开关切 side 时，对每个 `ReplayPlayer` 调用
+  `setFilter(side, 该图固定 rtype)`（合并图固定 Pistol，买局图固定 Buy）。
+- `RTYPES = ["Pistol","Full","Eco"]` 常量移除/改写为新布局逻辑。
+
+## 影响文件
+
+| 文件 | 改动 |
+|------|------|
+| `server/config.py` | 新增 `EQ_BUY_MIN = 2000` |
+| `server/parse.py` | `classify_rounds` 个人判定 + 丢弃逻辑；3 个下游过滤条件加 `and r["rtype"]` |
+| `server/combat.py` | AWP 改持狙占比（inventory 检测 + 回退）；返回值与聚合调整 |
+| `server/templates/index.html` | header 加统一 CT/T 开关；main 加 `#pistol` 区 |
+| `server/static/app.js` | 新布局：合并手枪图、单买局图、统一开关接线；移除三宫格/每卡 tabs |
+| `server/static/replay.js` | per-round 颜色覆盖 |
+| `server/player_json.py` | 无需改（rtype 自然变 Pistol/Buy） |
+
+## 验证
+
+- 单元/集成：对 fixture demo `demos_analysis/g161-...de_mirage.dem` 验证
+  `classify_rounds` 只产出 Pistol/Buy/None，个人判定生效；`parse_combat_stats` 持狙占比对已知
+  AWPer（如 76561199755381652，26 次 AWP 击杀）> 0。
+- 视觉：重启 `web_server.py`，跑一次真实扫描或用 `replay_test.html`，确认顶部统一开关切换全部
+  canvas、手枪合并图多色叠加、每卡单买局图正常。
+
+## 不做（YAGNI）
+
+- 不恢复 Force Buy 档。
+- 不做按回合逐个开关（每图整体叠加即可）。
+- 不改下载/管线/5E 抓取逻辑。