Harden STOM RL evidence dashboard and NO-GO research guardrails

.gitignore

@@ -7,11 +7,17 @@ __pycache__/
 build/
 develop-eggs/
 dist/
+# [REV-7] Vite 빌드 산출물은 prebuilt 정책으로 commit 한다 (위 dist/ 규칙의 예외)
+!webui/static/v2/dist/
+!webui/static/v2/dist/**
 downloads/
 eggs/
 .eggs/
 lib/
 lib64/
+# Python lib/ 규칙의 예외 — Svelte 소스 디렉터리는 commit 한다
+!webui/v2_src/src/lib/
+!webui/v2_src/src/lib/**
 parts/
 sdist/
 var/
@@ -48,16 +54,38 @@ Desktop.ini
 *.parquet
 *.h5
 *.hdf5
+_database/
+finetune_csv/data/stom_*.csv
+finetune_csv/data/stom_*.json
+webui/stom_predictions/*.csv
+webui/stom_predictions/*.json
+webui/stom_predictions/
+webui/qlib_backtests/*.json
+webui/qlib_backtests/*.csv
+webui/rl_runs/
+finetune/qlib_exports/
+finetune/qlib_backtests/
+finetune/outputs/
+finetune/data/processed_datasets/
+finetune/data/stom_qlib*/
+qlib_data/
 
 # Model files (large files)
 *.pth
 *.pt
 *.ckpt
 *.bin
+finetune_csv/finetuned/
 
 # Logs
 *.log
 logs/
+.omx/
+# Local agent/workflow/session state
+.codegraph/
+.gjc/
+.omc/
+.omo/
 
 # Environment
 .env

.omc/skill-candidates.md

@@ -0,0 +1,135 @@
+# Kronos v2 — 스킬 후보 (session-wrap 2026-05-18 추출)
+
+본 세션에서 7개 탭에 걸쳐 반복적으로 등장한 패턴들. 추후 `/oh-my-claudecode:skillify` 또는 `/oh-my-claudecode:learner` 로 재사용 스킬로 변환 가능.
+
+---
+
+## P1 · 5단 카드 헤더 패턴
+
+**등장**: 7개 탭 전부, ~20곳 이상
+
+```svelte
+<div class="card">
+  <div class="card-header">
+    <div>
+      <div class="card-eyebrow">{ENDPOINT_OR_PHASE}</div>
+      <div class="card-title">{HUMAN_TITLE}</div>
+    </div>
+    <span class="pill {STATUS_KIND}"><span class="dot"></span>{STATUS_LABEL}</span>
+  </div>
+  <!-- body -->
+</div>
+```
+
+5층 계층: card → header → eyebrow(작은 라벨) → title(굵은 제목) → 우측 status pill. 데이터 출처 표기(/api/...)는 eyebrow 에, 사람이 읽는 제목은 title 에, 현재 상태는 pill 에.
+
+---
+
+## P2 · KPI 메트릭 스트립 (.metric × N)
+
+**등장**: Live Training, History, System Health, STOM, Artifacts
+
+```svelte
+<section class="grid-1-1-1-1">
+  <div class="metric">
+    <div class="metric-head">
+      <span class="metric-label">{LABEL}</span>
+      <span class="delta {up|down|flat}">▼ {pct}%</span>
+    </div>
+    <div class="metric-value tnum">{value}<span class="metric-unit">{unit}</span></div>
+    <div class="metric-foot">{context_data}</div>
+  </div>
+  ...
+</section>
+```
+
+핵심: `tabular-nums` 로 숫자 정렬 + delta 화살표 색상 분기(▼=down=success / ▲=up=danger) + 미니 컨텍스트.
+
+---
+
+## P3 · Status Pill 4-way 분기
+
+**등장**: 모든 탭
+
+```ts
+function statusKind(s: string): 'success' | 'danger' | 'accent' | 'warn' | '' {
+  if (['completed','complete','success','done'].includes(s)) return 'success';
+  if (['failed','error'].includes(s)) return 'danger';
+  if (['running','active'].includes(s)) return 'accent';
+  if (['waiting','pending'].includes(s)) return 'warn';
+  return '';
+}
+```
+
+색상 매핑 단일 owner 함수 → CSS `.pill.success / .danger / .accent / .warn` 와 자동 매칭.
+
+---
+
+## P4 · ECharts theme-aware palette 추출
+
+**등장**: W3, W5, SystemHealth, Forecast
+
+```ts
+let palette = $derived.by(() => {
+  void currentTheme;  // ← theme store 의존성 강제 등록
+  if (typeof window === 'undefined') return null;
+  const cs = getComputedStyle(document.documentElement);
+  return {
+    accent: cs.getPropertyValue('--accent').trim(),
+    grid: cs.getPropertyValue('--border-faint').trim(),
+    text: cs.getPropertyValue('--fg').trim(),
+    // ...
+  };
+});
+```
+
+light/dark 전환 시 CSS 변수가 자동으로 바뀌고 `theme` store 구독으로 derived 가 재계산 → ECharts option 이 새 palette 로 setOption 호출됨.
+
+---
+
+## P5 · Svelte 5 stores + subscribe 패턴
+
+**등장**: 모든 컴포넌트
+
+```svelte
+<script lang="ts">
+  import { trainingStatus } from '$lib/stores';
+
+  let status = $state<any>(null);
+  trainingStatus.subscribe((v) => (status = v));
+
+  let latest = $derived(status?.latest_stage ?? {});
+  let pct = $derived(latest?.overall_percent ?? 0);
+</script>
+```
+
+핵심: `$state + subscribe + $derived` 3단 체인으로 store 데이터를 reactive 하게 사용. App.svelte 가 polling 시작 → stores 갱신 → 모든 컴포넌트가 자동 업데이트.
+
+---
+
+## P6 · data-active 속성 기반 토글
+
+**등장**: Sidebar nav, tabs, seg buttons, file rows
+
+```svelte
+<button
+  data-active={current === id ? 'true' : 'false'}
+  onclick={() => current = id}
+>...</button>
+```
+
+```css
+.nav-item[data-active="true"] { background: var(--accent-soft); border-color: var(--accent); }
+.tabs button[data-active="true"] { border-bottom: 2px solid var(--accent); color: var(--accent-strong); }
+```
+
+class 분기 대신 data attribute 로 토글 — CSS attribute selector 가 더 가벼우면서 JS 코드 단순화.
+
+---
+
+## 변환 우선순위
+
+가장 가치 큰 순서:
+1. **P1 (5단 카드 헤더)** — 새 탭 추가 시 boilerplate 자동화
+2. **P4 (ECharts palette 추출)** — 다른 차트 라이브러리 도입 시 동일 패턴 재사용
+3. **P3 (Status pill 분기)** — 다른 도메인의 status 표시 통일

AGENTS.md

@@ -0,0 +1,98 @@
+# Kronos Project Knowledge Base
+
+**Generated:** 2026-06-03 KST
+**Last reviewed:** 2026-06-03 KST, init-deep update mode
+**Branch observed:** `feature/stom-rl-lab`
+**Commit observed:** `943222b`
+
+## Overview
+
+Kronos is now a combined research/operations repo: core Kronos model code, STOM
+1-second/tick data pipelines, rule/RL trading research, and the official Flask +
+Svelte dashboard for inspection. Treat it as an experimental trading research
+platform, not as a live-trading product.
+
+## Current Direction
+
+| Area | Current stance | Reason |
+|---|---|---|
+| `ts_imb` opening gap-up rule | Main research baseline | Prior docs show the strongest useful curve here; it is a RULE strategy, not RL. |
+| Skip-gate / state-exit gates | Deprioritize unless new hypothesis is preregistered | Full-universe docs report `NO-GO`. |
+| Plain PPO/DQN RL | Do not present as usable | Existing PPO/DQN/orderbook runs are `NO-GO` or research-only. |
+| Orderbook RL | Keep as isolated experiment/falsification tool | Useful for testing action design and dashboard comparison, not live readiness. |
+| Dashboard | Continue as evidence viewer | It should expose failure, baselines, costs, and split metadata clearly. |
+
+## Structure
+
+```text
+Kronos/
++-- model/             # Kronos model/tokenizer implementation
++-- finetune/          # STOM/Qlib export, training, evaluation CLIs
++-- stom_rl/           # STOM rule/RL experiments, gates, backtests, readiness
++-- webui/             # Flask API and official dashboard adapters
++-- webui/v2_src/      # Svelte/Vite dashboard source (internal path name)
++-- tests/             # pytest coverage for model, STOM, web, dashboard
++-- docs/              # handoff, verdict, preregistration, result documents
++-- _database/         # local data; do not mutate casually
++-- .omx/artifacts/, webui/rl_runs/  # generated experiment artifacts
+```
+
+## Where To Look
+
+| Task | Location | Notes |
+|---|---|---|
+| RL/orderbook environment | `stom_rl/orderbook_rl_env.py` | Marketable-fill, orderbook feature environment. |
+| SB3 orderbook smoke | `stom_rl/orderbook_sb3_smoke.py` | Research-only DQN/PPO-style experiments. |
+| Gap-up rule baseline | `stom_rl/gap_up_backtest.py` | Main `ts_imb` rule reference. |
+| Skip/state gates | `stom_rl/skip_gate.py`, `stom_rl/state_exit_gate.py` | Existing full-universe `NO-GO` results. |
+| RL dashboard backend | `webui/rl_dashboard.py`, `webui/app.py` | Read-only artifact/API layer. |
+| RL dashboard frontend | `webui/v2_src/src/tabs/RLTradingTab.svelte` | Main RL/trading dashboard tab. |
+| Latest direction docs | `docs/stom_rl_resume_commit_2026-05-29.md`, `docs/stom_state_exit_result_2026-06-02.md` | Re-read before changing strategy direction. |
+
+## Trading Honesty Rules
+
+- Do **not** call the gap-up `ts_imb` curve "reinforcement learning". It is a
+  rule strategy unless a real RL policy produced it.
+- Do **not** claim live-trading readiness, profitability, or broker readiness.
+  Current work is local backtest/research/dashboard evidence only.
+- Primary cost assumption is 23bp round trip unless a test/document explicitly
+  states otherwise.
+- For high-frequency/opening work, prefer marketable-fill accounting where
+  possible (`buy@ask`, `sell@bid`) and label assumptions.
+- Preserve leading-zero stock codes. Never coerce codes like `000250` to int.
+- Treat local Korean DB columns and generated docs as UTF-8-sensitive.
+- Negative/shuffle controls and OOS splits are not optional for alpha claims.
+- If a model fails cost gate, baseline comparison, or drawdown gate, surface
+  `NO-GO` plainly in docs and UI.
+
+## Commands
+
+```powershell
+# Core RL/dashboard regression set
+py -3.11 -m pytest tests/test_stom_rl_dashboard_api.py tests/test_stom_rl_dashboard_tab.py tests/test_stom_rl_orderbook_env.py tests/test_stom_rl_orderbook_sb3.py -q
+
+# Broader STOM rule/gate checks
+py -3.11 -m pytest tests/test_stom_rl_gap_up_backtest.py tests/test_stom_rl_skip_gate.py tests/test_stom_rl_state_exit_gate.py tests/test_stom_rl_marketable_fill.py -q
+
+# Svelte dashboard build/check
+cd webui/v2_src
+npm run build
+```
+
+## Status Hygiene
+
+- Keep source/docs changes separate from generated outputs.
+- Treat `.omc/`, `.codegraph/`, `.omx/artifacts/`, `webui/rl_runs/`, and
+  frontend dist assets as generated/session state unless explicitly requested.
+- If committing later, review untracked files carefully before staging.
+- `webui/v2_src` expects Node 20 or 22 and npm 9+ per `package.json`.
+
+## Gotchas
+
+- `webui/rl_runs/`, `.omx/artifacts/`, `finetune/outputs/`, and large CSV/DB
+  directories are generated data. Use them as evidence; do not design from one
+  cherry-picked artifact.
+- `webui/v2_src` is a separate frontend project with its own package scripts.
+- `webui/app.py` is broad and central; API changes need targeted tests.
+- Existing docs intentionally use `NO-GO` heavily. Treat those as guardrails,
+  not as failure to hide.