chore(deps): update dependency uvicorn to v0.50.2

El panel Win Rate estuvo vacío los 81 días por diseño, no por bug: la API
emite win_rate: null hasta >=10 mercados resueltos (hubo 2) y Grafana no
renderiza null. Documentado en el informe final.

.gitea/workflows/ci.yml

@@ -1,9 +1,10 @@
 name: CI/CD
 
+# Decommission (2026-07): trigger switched from push:main to manual only, so
+# documentation commits don't rebuild/redeploy images. Run from the Gitea UI
+# (Actions → Run workflow) if a rebuild is ever needed again.
 on:
-  push:
-    branches:
-      - main
+  workflow_dispatch:
 
 env:
   REGISTRY: gitea.gitea.svc.cluster.local:3000
@@ -178,7 +179,8 @@ jobs:
           # keep their current (still existing) tag in the registry.
           if [ "${{ steps.changes.outputs.build_bot }}" = "true" ]; then
             sed -i "s|image: .*polymarket-bot[^-].*|image: git.chemavx.xyz/chemavx/polymarket-bot:${TAG}|g" \
-              polymarket-bot/deployment-bot.yaml
+              polymarket-bot/deployment-bot.yaml \
+              polymarket-bot/cronjob-outcomes.yaml
           fi
           if [ "${{ steps.changes.outputs.build_api }}" = "true" ]; then
             sed -i "s|image: .*polymarket-bot-api.*|image: git.chemavx.xyz/chemavx/polymarket-bot-api:${TAG}|g" \

README.md

@@ -1,7 +1,118 @@
 # polymarket-bot
 
-Bot de paper-trading para Polymarket con estrategia bayesiana, API FastAPI y
-dashboard React. Corre en k3s vía GitOps (Gitea Actions → registry → ArgoCD).
+Bot de **paper trading** para mercados de predicción de Polymarket: estrategia
+bayesiana sobre el precio de mercado como prior, señales externas en log-odds,
+gates de riesgo en cascada y un motor de replay determinista para calibración
+offline. API FastAPI + dashboard React. Desplegado en k3s vía GitOps
+(Gitea Actions → registry → ArgoCD), con PostgreSQL, secrets por Infisical y
+observabilidad completa (Grafana, uptime-kuma, Telegram).
+
+> **📦 Estado del proyecto: ARCHIVADO (julio 2026).**
+> El 26 de mayo de 2026 la DGOJ [ordenó el bloqueo cautelar de Polymarket y
+> Kalshi en España](https://www.ordenacionjuego.es/novedades/dgoj-abre-expediente-sancionador-plataformas-polymarket-kalshi-ordena-bloqueo-sus-webs)
+> por operar sin licencia de juego ([CoinDesk](https://www.coindesk.com/policy/2026/05/26/spain-joins-growing-list-of-countries-shutting-out-polymarket-and-kalshi)).
+> Con el bloqueo a nivel de ISP la fuente de datos primaria desaparece, así que
+> el proyecto se jubila de forma ordenada: datos respaldados y verificados,
+> resultados documentados en el [informe final](docs/informe-final.md), e
+> infraestructura apagada vía GitOps. El cierre es parte de la historia del
+> proyecto, no un abandono — los módulos centrales son agnósticos de la fuente
+> y el pivote natural es Manifold Markets (API abierta, play-money, sin
+> fricción regulatoria).
+
+## El problema y la solución
+
+Los mercados de predicción publican probabilidades implícitas en el precio.
+La hipótesis: en mercados de bajo volumen hay ineficiencias detectables
+combinando el precio con señales externas (noticias, sentimiento cripto,
+mercados espejo en otras plataformas). El bot la pone a prueba **sin arriesgar
+dinero**, con la disciplina de un sistema real:
+
+1. **Prior desde el precio** de Polymarket (mid del order book CLOB).
+2. **Ajuste bayesiano en log-odds** con señales independientes: GNews (con
+   presupuesto de 5 consultas/día y *guardrail* que limita el ajuste a
+   prior±0,25), Fear&Greed, momentum BTC, dominancia BTC, y Manifold Markets
+   como señal cruzada (en modo observacional tras detectarse un bug de
+   inversión — auditado en 165k filas).
+3. **Gates en cascada** antes de operar: prior extremo (<0,08 / >0,92), edge
+   bruto mínimo por régimen de volatilidad, edge neto positivo tras comisión
+   (2%) y spread, confianza ≥0,55, conflicto de familias, guard de reentrada.
+4. **Agrupación por familias de mercados**: mercados hermanos del mismo evento
+   (candidatos de una misma elección, umbrales de un mismo precio) comparten
+   `family_key`; solo se mantiene la posición de mayor edge para no apilar la
+   misma apuesta con dos nombres.
+5. **Sizing por Kelly fraccionado** con techo por posición sobre un bankroll
+   simulado de $10.000.
+
+Resultado honesto en 81 días: 12 trades (n estadísticamente irrelevante, y el
+informe lo subraya), realized **+$247,78**, y — más interesante — **~223.000
+evaluaciones archivadas con su decisión reproducible**, porque el sistema
+prefirió no operar antes que operar sin ventaja medible.
+Detalle completo: [docs/informe-final.md](docs/informe-final.md).
+
+## Arquitectura
+
+```mermaid
+flowchart LR
+    subgraph ext["Fuentes externas"]
+        PM["Polymarket<br/>CLOB + Gamma API"]
+        GN["GNews"]
+        MF["Manifold Markets"]
+        CG["CoinGecko /<br/>alternative.me"]
+    end
+
+    subgraph k3s["k3s · namespace polymarket-bot"]
+        BOT["bot (ciclo ~64s)<br/>estrategia bayesiana + gates"]
+        PG[("PostgreSQL 16<br/>trades · signals · replay")]
+        API["api (FastAPI :8000)"]
+        DASH["dashboard (React/nginx)"]
+        CJ1["CronJob metrics-retention 00:10"]
+        CJ2["CronJob outcomes-joiner 00:30"]
+    end
+
+    subgraph obs["Observabilidad"]
+        GRAF["Grafana"]
+        UK["uptime-kuma"]
+        TG["Telegram"]
+    end
+
+    PM --> BOT
+    GN --> BOT
+    MF --> BOT
+    CG --> BOT
+    BOT --> PG
+    PG --> API
+    API --> DASH
+    API --> GRAF
+    UK --> DASH
+    BOT --> TG
+    PM --> CJ2
+    CJ2 --> PG
+    CJ1 --> PG
+
+    subgraph ci["GitOps"]
+        GA["Gitea Actions"] --> REG["registry"] --> ARGO["ArgoCD<br/>prune + selfHeal"]
+    end
+    ARGO -.-> k3s
+```
+
+## Decisiones de diseño
+
+- **Paper trading por diseño, no como demo**: el bot exigía ≥10 mercados
+  resueltos y ≥30 días antes de considerarse promocionable a dinero real, y
+  reportaba `n/a (insufficient_sample)` en vez de métricas infladas.
+- **Todo skip es un dato**: cada evaluación archiva prior, estimación, edge
+  bruto/neto, descomposición por feature en log-odds y el gate exacto que la
+  bloqueó (~55k filas/día en `signals`).
+- **Replay determinista** (julio 2026): inyección de reloj + re-ejecución de
+  ciclos archivados con 22.697/22.697 decisiones idénticas; joiner diario de
+  resoluciones UMA para calibrar sobre *todas* las evaluaciones, no solo los
+  trades.
+- **Los errores se cierran y se documentan**: el bug de inversión de Manifold
+  y los conflictos de familia cerraron 5 posiciones en abril; la señal pasó a
+  observacional-only con auditoría completa en vez de eliminarse.
+- **GitOps estricto**: nunca `kubectl` para cambios persistentes (ArgoCD con
+  `selfHeal` revierte cualquier parche manual); secrets fuera de git vía
+  Infisical operator; smoke test PostSync con notificación a Telegram.
 
 ## Componentes
 
@@ -11,10 +122,19 @@ dashboard React. Corre en k3s vía GitOps (Gitea Actions → registry → ArgoCD
 | api | `api/` (+ `bot/` como librería) | `polymarket-bot-api` | `uvicorn api.main:app` |
 | dashboard | `dashboard/` | `polymarket-bot-dashboard` | nginx estático |
 
-Dashboard: https://polymarket.chemavx.xyz
+Dashboard (hasta el apagado): https://polymarket.chemavx.xyz
+
+## Stack
+
+Python 3 (asyncio) · FastAPI · React + Vite · PostgreSQL 16 · k3s · ArgoCD ·
+Gitea Actions + BuildKit · Infisical (secrets) · Grafana + Prometheus ·
+uptime-kuma · Telegram Bot API.
 
 ## CI/CD
 
+> ⚠️ **Decomisión**: el trigger automático `push:main` se desactivó en la Fase 1
+> del archivado; el workflow queda solo bajo `workflow_dispatch` manual.
+
 `.gitea/workflows/ci.yml` construye **solo las imágenes cuyos ficheros
 cambiaron** en el push (diff contra `github.event.before`):
 
@@ -33,3 +153,17 @@ sincroniza vía webhook en segundos.
 ```bash
 python3 -m pytest tests/ -q
 ```
+
+## Datos y backup
+
+Base de datos respaldada y verificada (2026-07-06): `pg_dump -Fc` + exports
+CSV.gz de las tablas de señales y auditoría, con checksums y restore de prueba
+(recuentos idénticos 11/11 tablas). Ubicación: almacenamiento de backups del
+clúster con sync offsite. El dump final se toma en la fase de apagado.
+
+## Documentación
+
+- [Informe final de paper trading](docs/informe-final.md) — resultados,
+  pipeline de datos y estado de la BD al cierre.
+- `docs/pivot-manifold.md` — notas de diseño del posible pivote (Fase 5,
+  pendiente).

bot/data/db.py

@@ -747,6 +747,154 @@ class Database:
         except (ValueError, IndexError):
             return 0
 
+    # ── Replay R1: replay core ───────────────────────────────────────────────
+
+    async def get_replay_cycles(self, from_ts, to_ts) -> list:
+        """Return the cycle_ts values with archived decisions in [from_ts, to_ts)."""
+        async with self._pool.acquire() as conn:
+            rows = await conn.fetch("""
+                SELECT DISTINCT cycle_ts FROM signals
+                WHERE cycle_ts >= $1 AND cycle_ts < $2
+                ORDER BY cycle_ts
+            """, from_ts, to_ts)
+        return [r["cycle_ts"] for r in rows]
+
+    async def get_ext_snapshot(self, cycle_ts) -> Optional[dict]:
+        """Return one cycle's ExternalSignals snapshot, or None if missing."""
+        async with self._pool.acquire() as conn:
+            row = await conn.fetchrow(
+                "SELECT * FROM ext_snapshots WHERE cycle_ts = $1", cycle_ts
+            )
+        return dict(row) if row else None
+
+    async def get_cycle_signal_rows(self, cycle_ts) -> list[dict]:
+        """Return one cycle's archived decision rows in original evaluation
+        order (id = insertion order = the order main.py evaluated them)."""
+        async with self._pool.acquire() as conn:
+            rows = await conn.fetch(
+                "SELECT * FROM signals WHERE cycle_ts = $1 ORDER BY id", cycle_ts
+            )
+        return [dict(r) for r in rows]
+
+    async def get_markets_by_ids(self, market_ids: list[str]) -> dict[str, dict]:
+        """Return market metadata rows keyed by id (for Market reconstruction)."""
+        if not market_ids:
+            return {}
+        async with self._pool.acquire() as conn:
+            rows = await conn.fetch(
+                "SELECT * FROM markets WHERE id = ANY($1::text[])", market_ids
+            )
+        return {r["id"]: dict(r) for r in rows}
+
+    async def save_replay_run(self, run: dict) -> None:
+        async with self._pool.acquire() as conn:
+            await conn.execute("""
+                INSERT INTO replay_runs (
+                    run_id, git_sha, config_hash, config_json,
+                    from_ts, to_ts, cycles, decisions, matched, mismatched, note
+                ) VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11)
+            """,
+                run["run_id"], run["git_sha"], run["config_hash"],
+                run["config_json"], run["from_ts"], run["to_ts"],
+                run["cycles"], run["decisions"], run["matched"],
+                run["mismatched"], run["note"],
+            )
+
+    async def save_replay_decisions(self, run_id: str, decisions: list[dict]) -> None:
+        if not decisions:
+            return
+        rows = [
+            (
+                run_id, d["cycle_ts"], d["market_id"],
+                d["skip_reason"], d["prior_prob"], d["estimated_prob"],
+                d["raw_final_prob"], d["edge_gross"], d["edge_net"],
+                d["regime_min_edge"], d["days_to_resolution"],
+                d["confidence"], d["direction"], d["would_trade"],
+                d["recorded_skip_reason"], d["matched"], d["mismatch_field"],
+            )
+            for d in decisions
+        ]
+        async with self._pool.acquire() as conn:
+            await conn.executemany("""
+                INSERT INTO replay_decisions (
+                    run_id, cycle_ts, market_id,
+                    skip_reason, prior_prob, estimated_prob,
+                    raw_final_prob, edge_gross, edge_net,
+                    regime_min_edge, days_to_resolution,
+                    confidence, direction, would_trade,
+                    recorded_skip_reason, matched, mismatch_field
+                ) VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15,$16,$17)
+            """, rows)
+
+    # ── Replay R2: outcomes + calibration metrics ────────────────────────────
+
+    async def get_unresolved_archived_market_ids(self) -> list[str]:
+        """Archived markets (present in signals) with no stored outcome yet."""
+        async with self._pool.acquire() as conn:
+            rows = await conn.fetch("""
+                SELECT DISTINCT s.market_id FROM signals s
+                LEFT JOIN market_outcomes o ON o.market_id = s.market_id
+                WHERE o.market_id IS NULL
+                ORDER BY s.market_id
+            """)
+        return [r["market_id"] for r in rows]
+
+    async def upsert_market_outcome(
+        self, market_id: str, outcome: float, resolved_at
+    ) -> None:
+        async with self._pool.acquire() as conn:
+            await conn.execute("""
+                INSERT INTO market_outcomes (market_id, outcome, resolved_at)
+                VALUES ($1, $2, $3)
+                ON CONFLICT (market_id) DO UPDATE
+                SET outcome = EXCLUDED.outcome,
+                    resolved_at = EXCLUDED.resolved_at,
+                    fetched_at = NOW()
+            """, market_id, outcome, resolved_at)
+
+    async def get_outcome_coverage(self) -> dict:
+        """How much of the archive is scorable: resolved vs archived markets."""
+        async with self._pool.acquire() as conn:
+            row = await conn.fetchrow("""
+                SELECT
+                    (SELECT COUNT(DISTINCT market_id) FROM signals)  AS archived,
+                    (SELECT COUNT(*) FROM market_outcomes
+                     WHERE market_id IN (SELECT DISTINCT market_id FROM signals)
+                    ) AS resolved
+            """)
+        return dict(row)
+
+    async def get_calibration_rows(self, run_id: Optional[str] = None) -> list[dict]:
+        """Every archived evaluation with a full estimate AND a known outcome.
+
+        run_id None scores the R0 archive (signals); a run_id scores that
+        replay run's re-estimates instead (counterfactual calibration).
+        Rows without estimated_prob (skipped before estimation: prior_extreme,
+        unsupported, family, no_signals) carry no model prediction to score.
+        """
+        async with self._pool.acquire() as conn:
+            if run_id is None:
+                rows = await conn.fetch("""
+                    SELECT s.market_id, s.category,
+                           s.estimated_prob, s.prior_prob, o.outcome
+                    FROM signals s
+                    JOIN market_outcomes o ON o.market_id = s.market_id
+                    WHERE s.estimated_prob IS NOT NULL
+                      AND s.prior_prob IS NOT NULL
+                """)
+            else:
+                rows = await conn.fetch("""
+                    SELECT d.market_id, m.category,
+                           d.estimated_prob, d.prior_prob, o.outcome
+                    FROM replay_decisions d
+                    JOIN market_outcomes o ON o.market_id = d.market_id
+                    LEFT JOIN markets m ON m.id = d.market_id
+                    WHERE d.run_id = $1
+                      AND d.estimated_prob IS NOT NULL
+                      AND d.prior_prob IS NOT NULL
+                """, run_id)
+        return [dict(r) for r in rows]
+
     async def mark_manifold_audit_used(self, audit_id: str) -> None:
         async with self._pool.acquire() as conn:
             await conn.execute(

bot/data/schema.sql

@@ -370,3 +370,85 @@ CREATE TABLE IF NOT EXISTS ext_snapshots (
     total_market_cap_change DOUBLE PRECISION,
     valid                   BOOLEAN
 );
+
+-- ─────────────────────────────────────────────────────────────────────────────
+-- Replay R1: replay core — re-execute evaluate() over the R0 archive
+--
+-- A replay run reads cycles from signals + ext_snapshots + markets, rebuilds
+-- the exact inputs (including archived news_sentiment — GNews is never called),
+-- re-runs BayesianStrategy.evaluate() with the archived cycle_ts as clock, and
+-- writes one replay_decisions row per (cycle, market).
+--
+-- replay_runs tags every run with the code (git_sha) and strategy constants
+-- (config_hash) that produced it: two runs over the same window with different
+-- config_hash values are a counterfactual comparison; same config_hash against
+-- the recorded rows is a determinism check (mismatches should be 0, modulo
+-- day-boundary crossings between cycle_ts and the original wall-clock).
+--
+-- matched: replayed decision equals the recorded one (skip_reason, probs,
+--   confidence, direction).  NULL when not comparable — e.g. reentry_guard
+--   rows, recorded outside evaluate() with no decision fields to compare;
+--   the replay still re-evaluates them, which is extra calibration data.
+-- mismatch_field: first field that differed, for triage.
+-- ─────────────────────────────────────────────────────────────────────────────
+CREATE TABLE IF NOT EXISTS replay_runs (
+    run_id      TEXT PRIMARY KEY,
+    created_at  TIMESTAMPTZ DEFAULT NOW(),
+    git_sha     TEXT,
+    config_hash TEXT,
+    config_json TEXT,
+    from_ts     TIMESTAMPTZ,
+    to_ts       TIMESTAMPTZ,
+    cycles      INTEGER,
+    decisions   INTEGER,
+    matched     INTEGER,
+    mismatched  INTEGER,
+    note        TEXT
+);
+
+CREATE TABLE IF NOT EXISTS replay_decisions (
+    id                   SERIAL PRIMARY KEY,
+    run_id               TEXT NOT NULL,
+    cycle_ts             TIMESTAMPTZ NOT NULL,
+    market_id            TEXT NOT NULL,
+    -- replayed outputs (same semantics as the signals columns)
+    skip_reason          TEXT,
+    prior_prob           DOUBLE PRECISION,
+    estimated_prob       DOUBLE PRECISION,
+    raw_final_prob       DOUBLE PRECISION,
+    edge_gross           DOUBLE PRECISION,
+    edge_net             DOUBLE PRECISION,
+    regime_min_edge      DOUBLE PRECISION,
+    days_to_resolution   INTEGER,
+    confidence           DOUBLE PRECISION,
+    direction            TEXT,
+    would_trade          BOOLEAN,
+    -- fidelity vs the recorded signals row
+    recorded_skip_reason TEXT,
+    matched              BOOLEAN,
+    mismatch_field       TEXT
+);
+
+CREATE INDEX IF NOT EXISTS idx_replay_decisions_run ON replay_decisions(run_id);
+CREATE INDEX IF NOT EXISTS idx_replay_decisions_mkt ON replay_decisions(market_id);
+
+-- ─────────────────────────────────────────────────────────────────────────────
+-- Replay R2: outcomes + calibration metrics
+--
+-- One row per resolved market, fetched from the Gamma API via
+-- get_market_resolution() (UMA-final only: a market closed but still in
+-- proposal/dispute is not stored).  outcome is the final YES price:
+-- 1.0 = YES won, 0.0 = NO won.
+--
+-- Joining signals (or replay_decisions) to market_outcomes scores every
+-- archived estimate against reality — Brier / log-loss of estimated_prob
+-- benchmarked against the market price (prior_prob) on the same rows,
+-- answering "does the model add value over the market?" across ALL
+-- evaluations, not just executed trades.
+-- ─────────────────────────────────────────────────────────────────────────────
+CREATE TABLE IF NOT EXISTS market_outcomes (
+    market_id   TEXT PRIMARY KEY,
+    outcome     DOUBLE PRECISION NOT NULL,
+    resolved_at TIMESTAMPTZ,
+    fetched_at  TIMESTAMPTZ DEFAULT NOW()
+);

bot/outcomes.py

@@ -0,0 +1,208 @@
+"""
+Replay R2 — outcomes + calibration metrics.
+
+Two phases, one CLI:
+
+1. Fetch: for every archived market (present in `signals`) without a stored
+   outcome, ask the Gamma API via PolymarketClient.get_market_resolution()
+   — the same UMA-finality gate the trading loop uses to settle positions.
+   Definitive resolutions are upserted into `market_outcomes`; open, disputed
+   or ambiguous markets are simply retried on the next invocation.  There is
+   no data-loss urgency here (unlike the R0 recorder): Gamma reports past
+   resolutions at any time, so running this lazily loses nothing.
+
+2. Score: join archived estimates to outcomes and compute Brier / log-loss of
+   estimated_prob, benchmarked against the market price (prior_prob) on the
+   same rows.  This scores ALL evaluations with a full estimate — the sample
+   multiplier the phase plan calls for — not just executed trades.  With
+   --run-id it scores a replay run's re-estimates instead (counterfactual
+   calibration: did config X predict better than the market?).
+
+Reading the numbers: lower is better for both metrics; model < prior means
+the model added information over the market price.  Micro averages weight
+every evaluation equally, so long-lived markets (~1 evaluation/min while in
+the universe) dominate; macro averages score each market once (mean of its
+evaluations) and answer the same question per market.  Evaluations of one
+market minutes apart are highly autocorrelated — n_evaluations overstates
+the effective sample size, n_markets is the honest one.
+
+CLI:
+    python -m bot.outcomes                  # fetch new outcomes, then score archive
+    python -m bot.outcomes --fetch-only
+    python -m bot.outcomes --metrics-only
+    python -m bot.outcomes --run-id UUID    # score a replay run (implies no fetch)
+"""
+import argparse
+import asyncio
+import logging
+import math
+from collections import defaultdict
+from typing import Optional
+
+from bot.data.db import Database
+from bot.data.polymarket import PolymarketClient
+
+log = logging.getLogger(__name__)
+
+# Clip probabilities before log() so a (theoretical) hard 0/1 estimate on a
+# wrong outcome scores ~20.7 nats instead of infinity poisoning the mean.
+LOGLOSS_EPS = 1e-9
+
+
+async def fetch_outcomes(poly, market_ids: list[str]) -> list[dict]:
+    """Resolve archived markets against Gamma; returns only definitive ones.
+
+    Sequential on purpose: ~50 markets per invocation, and the Gamma API has
+    no bulk endpoint.  get_market_resolution() already returns None on API
+    errors and resolved=False on open/disputed/ambiguous markets.
+    """
+    resolved = []
+    for market_id in market_ids:
+        res = await poly.get_market_resolution(market_id)
+        if res is None or not res.resolved or res.resolution is None:
+            continue
+        resolved.append({
+            "market_id": market_id,
+            "outcome": res.resolution,
+            "resolved_at": res.resolved_at,
+        })
+    return resolved
+
+
+def _logloss(p: float, outcome: float) -> float:
+    p = min(max(p, LOGLOSS_EPS), 1.0 - LOGLOSS_EPS)
+    return -math.log(p) if outcome == 1.0 else -math.log(1.0 - p)
+
+
+def compute_calibration(rows: list[dict]) -> Optional[dict]:
+    """Score estimated_prob vs prior_prob against outcomes; None if no rows.
+
+    rows: dicts with market_id, category, estimated_prob, prior_prob, outcome.
+    Pure function — the CLI feeds it DB rows, tests feed it literals.
+    """
+    if not rows:
+        return None
+
+    n = len(rows)
+    brier_model = sum((r["estimated_prob"] - r["outcome"]) ** 2 for r in rows) / n
+    brier_prior = sum((r["prior_prob"] - r["outcome"]) ** 2 for r in rows) / n
+    logloss_model = sum(_logloss(r["estimated_prob"], r["outcome"]) for r in rows) / n
+    logloss_prior = sum(_logloss(r["prior_prob"], r["outcome"]) for r in rows) / n
+
+    by_market: dict[str, list[dict]] = defaultdict(list)
+    for r in rows:
+        by_market[r["market_id"]].append(r)
+    market_briers = [
+        (
+            sum((r["estimated_prob"] - r["outcome"]) ** 2 for r in mrows) / len(mrows),
+            sum((r["prior_prob"] - r["outcome"]) ** 2 for r in mrows) / len(mrows),
+        )
+        for mrows in by_market.values()
+    ]
+    brier_model_macro = sum(b[0] for b in market_briers) / len(market_briers)
+    brier_prior_macro = sum(b[1] for b in market_briers) / len(market_briers)
+
+    by_category: dict[str, list[dict]] = defaultdict(list)
+    for r in rows:
+        by_category[r["category"] or "unknown"].append(r)
+    per_category = {
+        cat: {
+            "n": len(crows),
+            "markets": len({r["market_id"] for r in crows}),
+            "brier_model": sum((r["estimated_prob"] - r["outcome"]) ** 2
+                               for r in crows) / len(crows),
+            "brier_prior": sum((r["prior_prob"] - r["outcome"]) ** 2
+                               for r in crows) / len(crows),
+        }
+        for cat, crows in sorted(by_category.items())
+    }
+
+    return {
+        "n_evaluations": n,
+        "n_markets": len(by_market),
+        "brier_model": brier_model,
+        "brier_prior": brier_prior,
+        "brier_model_macro": brier_model_macro,
+        "brier_prior_macro": brier_prior_macro,
+        "logloss_model": logloss_model,
+        "logloss_prior": logloss_prior,
+        "per_category": per_category,
+    }
+
+
+def print_report(metrics: Optional[dict], source: str) -> None:
+    if metrics is None:
+        print(f"calibration : no scorable rows yet for {source} "
+              "(no archived estimate has a resolved outcome)")
+        return
+    print(f"calibration : {source} — {metrics['n_evaluations']} evaluations, "
+          f"{metrics['n_markets']} markets")
+    print(f"{'':14s}{'model':>10s}{'market':>10s}{'delta':>10s}")
+    for label, m_key, p_key in (
+        ("Brier micro", "brier_model", "brier_prior"),
+        ("Brier macro", "brier_model_macro", "brier_prior_macro"),
+        ("logloss micro", "logloss_model", "logloss_prior"),
+    ):
+        m, p = metrics[m_key], metrics[p_key]
+        print(f"  {label:12s}{m:>10.4f}{p:>10.4f}{m - p:>+10.4f}")
+    print("  (delta < 0 = model beats the market price)")
+    for cat, c in metrics["per_category"].items():
+        print(f"  {cat:12s}n={c['n']:<6d} markets={c['markets']:<3d} "
+              f"brier model {c['brier_model']:.4f} vs market {c['brier_prior']:.4f}")
+
+
+async def _amain(args: argparse.Namespace) -> None:
+    db = Database()
+    await db.connect()
+    try:
+        if not args.metrics_only and args.run_id is None:
+            pending = await db.get_unresolved_archived_market_ids()
+            poly = PolymarketClient()
+            try:
+                resolved = await fetch_outcomes(poly, pending)
+            finally:
+                await poly.close()
+            for out in resolved:
+                await db.upsert_market_outcome(
+                    out["market_id"], out["outcome"], out["resolved_at"]
+                )
+            print(f"outcomes    : {len(resolved)} newly resolved "
+                  f"(of {len(pending)} pending markets checked)")
+
+        coverage = await db.get_outcome_coverage()
+        print(f"coverage    : {coverage['resolved']}/{coverage['archived']} "
+              "archived markets resolved")
+
+        if args.fetch_only:
+            return
+        rows = await db.get_calibration_rows(run_id=args.run_id)
+        source = f"replay run {args.run_id}" if args.run_id else "R0 archive"
+        print_report(compute_calibration(rows), source)
+    finally:
+        await db.disconnect()
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        prog="python -m bot.outcomes",
+        description="Fetch market resolutions and score archived estimates.",
+    )
+    parser.add_argument("--fetch-only", action="store_true",
+                        help="only fetch/store outcomes, skip metrics")
+    parser.add_argument("--metrics-only", action="store_true",
+                        help="skip the Gamma fetch, score what is stored")
+    parser.add_argument("--run-id", default=None,
+                        help="score a replay run's re-estimates instead of "
+                             "the R0 archive (implies --metrics-only)")
+    args = parser.parse_args()
+    if args.fetch_only and args.metrics_only:
+        parser.error("--fetch-only and --metrics-only are mutually exclusive")
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
+    )
+    asyncio.run(_amain(args))
+
+
+if __name__ == "__main__":
+    main()

bot/replay.py

@@ -0,0 +1,394 @@
+"""
+Replay R1 — replay core.
+
+Re-executes BayesianStrategy.evaluate() over the R0 archive (signals +
+ext_snapshots + markets) and stores the outcome in replay_runs /
+replay_decisions.
+
+Determinism contract: evaluate() is a pure function of
+(market, ext, occupied_families, as_of) plus the news client, so a replay
+rebuilds exactly those four inputs from the archive:
+
+  market   — metadata from `markets`, per-cycle price/volume from `signals`
+  ext      — the cycle's `ext_snapshots` row
+  families — a family-skipped row replays with its own family_key occupied;
+             every other row replays with no occupancy (the recorded
+             skip_reason already reflects the original portfolio state)
+  as_of    — the archived cycle_ts (clock injection, Replay R1)
+
+GNews is never called: ReplayNews feeds back the archived news_sentiment.
+The per-cycle query budget is bypassed (reset before every market) because
+the archived sentiment already encodes the budget's effect — a
+budget-skipped market was recorded with sentiment 0.0.
+
+Manifold and the DB are not wired into the replayed strategy (manifold=None,
+db=None): the signal is observational-only in production (feat_mfld_lo is
+always 0.0 in the archive), so the replay reproduces decisions without
+touching cooldowns or audit tables.  If MANIFOLD_SIGNAL_ENABLED is ever
+turned on, replayed decisions will diverge from recorded ones and the
+matched/mismatch_field columns will say so.
+
+Run tagging: every run stores the git sha and a hash of the strategy
+constants.  Same config_hash vs the archive = determinism check (expect 0
+mismatches, modulo UTC-day-boundary crossings between cycle_ts and the
+original wall-clock).  Different config_hash = counterfactual run.
+
+CLI:
+    python -m bot.replay --from 2026-07-02T00:00:00Z --to 2026-07-03 --note "..."
+"""
+import argparse
+import asyncio
+import hashlib
+import json
+import logging
+import os
+import subprocess
+import uuid
+from collections import Counter
+from datetime import datetime, timedelta, timezone
+from typing import Optional
+
+import bot.strategy.bayesian as bayesian
+from bot.data.db import Database
+from bot.data.external import ExternalSignals
+from bot.data.polymarket import Market
+from bot.strategy.bayesian import BayesianStrategy
+
+log = logging.getLogger(__name__)
+
+# Absolute float tolerance for recorded-vs-replayed comparison.  Archived
+# values are float8 (exact IEEE-754 round-trip of Python floats), so any real
+# divergence is far larger than this.
+FLOAT_TOL = 1e-9
+
+# Strategy constants that define a replay configuration.  Hashed into
+# replay_runs.config_hash; read from the module at call time so a
+# counterfactual run can monkeypatch them and be tagged distinctly.
+CONFIG_KEYS = (
+    "SPREAD_ESTIMATE",
+    "COMMISSION_RATE",
+    "MIN_CONFIDENCE",
+    "NEWS_LOGODDS_WEIGHT",
+    "MANIFOLD_LOGODDS_WEIGHT",
+    "MANIFOLD_SIGNAL_ENABLED",
+    "NEWS_GUARDRAIL_ENABLED",
+    "MAX_NEWS_ONLY_PROB_SHIFT",
+    "NEWS_MATERIAL_LOGODDS_THRESHOLD",
+    "MAX_NEWS_QUERIES_PER_CYCLE",
+)
+
+# Rows recorded outside evaluate() (via record_skip) carry no decision fields;
+# the replay still re-evaluates them for calibration but cannot compare.
+NON_COMPARABLE_SKIPS = {"reentry_guard"}
+
+
+def strategy_config() -> dict:
+    return {k: getattr(bayesian, k) for k in CONFIG_KEYS}
+
+
+def strategy_config_hash() -> str:
+    blob = json.dumps(strategy_config(), sort_keys=True)
+    return hashlib.sha256(blob.encode()).hexdigest()[:12]
+
+
+def _git_sha() -> str:
+    sha = os.getenv("GIT_SHA", "")
+    if sha:
+        return sha
+    try:
+        return subprocess.run(
+            ["git", "rev-parse", "--short", "HEAD"],
+            capture_output=True, text=True, timeout=5,
+        ).stdout.strip() or "unknown"
+    except (OSError, subprocess.SubprocessError):
+        return "unknown"
+
+
+class ReplayNews:
+    """NewsClient stand-in that feeds archived sentiment back into evaluate().
+
+    No HTTP, no cache: the engine sets `sentiment` to the archived value
+    before each evaluate() call.  Values below evaluate()'s 0.05 materiality
+    threshold were archived as 0.0, so the round-trip is exact.
+    """
+    enabled = True
+
+    def __init__(self) -> None:
+        self.sentiment: float = 0.0
+
+    async def get_sentiment(self, question: str) -> float:
+        return self.sentiment
+
+    def get_freshness(self, question: str) -> float:
+        return 1.0  # only used by gnews_priority(), which replay never calls
+
+
+def build_ext(snapshot: dict) -> ExternalSignals:
+    """Rebuild the ExternalSignals a cycle was evaluated against."""
+    return ExternalSignals(
+        btc_price=snapshot["btc_price"],
+        btc_change_24h=snapshot["btc_change_24h"],
+        eth_price=snapshot["eth_price"],
+        eth_change_24h=snapshot["eth_change_24h"],
+        btc_dominance=snapshot["btc_dominance"],
+        fear_greed_index=snapshot["fear_greed_index"],
+        fear_greed_label=snapshot["fear_greed_label"],
+        total_market_cap_change=snapshot["total_market_cap_change"],
+        valid=snapshot["valid"],
+    )
+
+
+def build_market(market_row: dict, signal_row: dict) -> Market:
+    """Rebuild a Market: metadata from `markets`, per-cycle state from `signals`.
+
+    Token ids are irrelevant to evaluate() and left empty; no_price is the
+    YES complement (evaluate() never reads it either).
+    """
+    yes_price = signal_row["polymarket_price"]
+    return Market(
+        id=market_row["id"],
+        condition_id=market_row["condition_id"] or "",
+        question=market_row["question"],
+        yes_token_id="",
+        no_token_id="",
+        yes_price=yes_price,
+        no_price=1.0 - yes_price,
+        volume_24h=signal_row["volume_24h"] or 0.0,
+        end_date=market_row["end_date"] or "",
+        active=True,
+        category=signal_row["category"] or (market_row["category"] or ""),
+    )
+
+
+def _compare(recorded: dict, replayed: dict) -> Optional[str]:
+    """Return the first field where replayed diverges from recorded, or None."""
+    if recorded["skip_reason"] != replayed["skip_reason"]:
+        return "skip_reason"
+    for field in ("prior_prob", "estimated_prob", "raw_final_prob",
+                  "edge_net", "confidence"):
+        a, b = recorded[field], replayed[field]
+        if a is None and b is None:
+            continue
+        if a is None or b is None or abs(a - b) > FLOAT_TOL:
+            return field
+    if recorded["direction"] != replayed["direction"]:
+        return "direction"
+    return None
+
+
+async def replay_cycle(
+    cycle_ts: datetime,
+    snapshot: dict,
+    signal_rows: list[dict],
+    market_rows: dict[str, dict],
+) -> list[dict]:
+    """Re-evaluate one archived cycle; returns one decision dict per row.
+
+    Pure with respect to the DB — everything it needs is passed in, so tests
+    can drive it with synthetic rows.
+    """
+    news = ReplayNews()
+    strategy = BayesianStrategy(news=news, manifold=None, db=None)
+    ext = build_ext(snapshot)
+    decisions: list[dict] = []
+
+    for row in signal_rows:
+        recorded_skip = row["skip_reason"]
+        decision = {
+            "cycle_ts": cycle_ts,
+            "market_id": row["market_id"],
+            "skip_reason": None,
+            "prior_prob": None,
+            "estimated_prob": None,
+            "raw_final_prob": None,
+            "edge_gross": None,
+            "edge_net": None,
+            "regime_min_edge": None,
+            "days_to_resolution": None,
+            "confidence": None,
+            "direction": None,
+            "would_trade": None,
+            "recorded_skip_reason": recorded_skip,
+            "matched": None,
+            "mismatch_field": None,
+        }
+
+        market_row = market_rows.get(row["market_id"])
+        if market_row is None:
+            # Should not happen (R0 upserts markets every cycle) — record the
+            # gap instead of crashing the run.
+            decision["matched"] = False
+            decision["mismatch_field"] = "market_missing"
+            decisions.append(decision)
+            continue
+
+        market = build_market(market_row, row)
+        # A family-skipped row replays against its own occupied family; all
+        # other rows replay unoccupied — their recorded skip_reason already
+        # reflects whatever portfolio state existed, and evaluate() checks
+        # the family gate before anything portfolio-dependent.
+        families = (
+            {row["family_key"]}
+            if recorded_skip == "family" and row["family_key"]
+            else set()
+        )
+        news.sentiment = row["news_sentiment"] or 0.0
+        # Bypass the per-cycle GNews budget: archived sentiment already
+        # encodes it (budget-skipped markets were recorded with 0.0).
+        strategy._news_queries_this_cycle = 0
+
+        signal = await strategy.evaluate(market, ext, families, as_of=cycle_ts)
+        rec = strategy.drain_cycle_records()[-1]
+
+        decision.update(
+            skip_reason=rec["skip_reason"],
+            prior_prob=rec["prior_prob"],
+            estimated_prob=rec["estimated_prob"],
+            raw_final_prob=rec["raw_final_prob"],
+            edge_gross=rec["edge_gross"],
+            edge_net=rec["edge_net"],
+            regime_min_edge=rec["regime_min_edge"],
+            days_to_resolution=rec["days_to_resolution"],
+            confidence=rec["confidence"],
+            direction=rec["direction"],
+            would_trade=signal is not None,
+        )
+        if recorded_skip in NON_COMPARABLE_SKIPS:
+            decision["matched"] = None  # re-evaluated for calibration only
+        else:
+            mismatch = _compare(row, rec)
+            decision["matched"] = mismatch is None
+            decision["mismatch_field"] = mismatch
+        decisions.append(decision)
+
+    return decisions
+
+
+async def run_replay(
+    db: Database,
+    from_ts: datetime,
+    to_ts: datetime,
+    note: str = "",
+    limit_cycles: Optional[int] = None,
+) -> dict:
+    """Replay every archived cycle in [from_ts, to_ts) and persist the run.
+
+    Returns the replay_runs row (plus a mismatch_fields Counter) for reporting.
+    """
+    run_id = str(uuid.uuid4())
+    cycles = await db.get_replay_cycles(from_ts, to_ts)
+    if limit_cycles:
+        cycles = cycles[:limit_cycles]
+
+    decisions_total = 0
+    matched = 0
+    mismatched = 0
+    mismatch_fields: Counter = Counter()
+    skipped_cycles = 0
+
+    for cycle_ts in cycles:
+        snapshot = await db.get_ext_snapshot(cycle_ts)
+        if snapshot is None:
+            skipped_cycles += 1
+            log.warning("Replay: no ext_snapshot for cycle %s — skipped", cycle_ts)
+            continue
+        signal_rows = await db.get_cycle_signal_rows(cycle_ts)
+        market_rows = await db.get_markets_by_ids(
+            [r["market_id"] for r in signal_rows]
+        )
+        decisions = await replay_cycle(cycle_ts, snapshot, signal_rows, market_rows)
+        await db.save_replay_decisions(run_id, decisions)
+
+        decisions_total += len(decisions)
+        for d in decisions:
+            if d["matched"] is True:
+                matched += 1
+            elif d["matched"] is False:
+                mismatched += 1
+                mismatch_fields[d["mismatch_field"]] += 1
+
+    run = {
+        "run_id": run_id,
+        "git_sha": _git_sha(),
+        "config_hash": strategy_config_hash(),
+        "config_json": json.dumps(strategy_config(), sort_keys=True),
+        "from_ts": from_ts,
+        "to_ts": to_ts,
+        "cycles": len(cycles) - skipped_cycles,
+        "decisions": decisions_total,
+        "matched": matched,
+        "mismatched": mismatched,
+        "note": note,
+    }
+    await db.save_replay_run(run)
+    run["mismatch_fields"] = dict(mismatch_fields)
+    run["skipped_cycles"] = skipped_cycles
+    return run
+
+
+def _parse_ts(value: str) -> datetime:
+    dt = datetime.fromisoformat(value.replace("Z", "+00:00"))
+    if dt.tzinfo is None:
+        dt = dt.replace(tzinfo=timezone.utc)
+    return dt
+
+
+async def _amain(args: argparse.Namespace) -> None:
+    db = Database()
+    await db.connect()
+    try:
+        run = await run_replay(
+            db,
+            from_ts=args.from_ts,
+            to_ts=args.to_ts,
+            note=args.note,
+            limit_cycles=args.limit_cycles,
+        )
+    finally:
+        await db.disconnect()
+
+    comparable = run["matched"] + run["mismatched"]
+    print(f"run_id      : {run['run_id']}")
+    print(f"git_sha     : {run['git_sha']}   config_hash: {run['config_hash']}")
+    print(f"window      : {run['from_ts'].isoformat()} → {run['to_ts'].isoformat()}")
+    print(f"cycles      : {run['cycles']} (skipped: {run['skipped_cycles']})")
+    print(f"decisions   : {run['decisions']} ({comparable} comparable)")
+    print(f"matched     : {run['matched']}")
+    print(f"mismatched  : {run['mismatched']}")
+    if run["mismatch_fields"]:
+        for field, count in sorted(run["mismatch_fields"].items(), key=lambda x: -x[1]):
+            print(f"  {field}: {count}")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        prog="python -m bot.replay",
+        description="Replay archived trading cycles through the current strategy.",
+    )
+    now = datetime.now(timezone.utc)
+    parser.add_argument(
+        "--from", dest="from_ts", type=_parse_ts,
+        default=now - timedelta(hours=24),
+        help="window start, ISO-8601 (default: 24h ago)",
+    )
+    parser.add_argument(
+        "--to", dest="to_ts", type=_parse_ts, default=now,
+        help="window end, ISO-8601, exclusive (default: now)",
+    )
+    parser.add_argument("--note", default="", help="free-text tag for replay_runs")
+    parser.add_argument(
+        "--limit-cycles", type=int, default=None,
+        help="replay at most N cycles (smoke runs)",
+    )
+    args = parser.parse_args()
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
+    )
+    # evaluate() logs one INFO line per market — thousands per replay window.
+    logging.getLogger("bot.strategy.bayesian").setLevel(logging.WARNING)
+    asyncio.run(_amain(args))
+
+
+if __name__ == "__main__":
+    main()

bot/strategy/bayesian.py

@@ -167,15 +167,22 @@ def _regime_min_edge(category: str, days_to_resolution: int) -> float:
     return 0.10  # tech, crypto/finance, events, default
 
 
-def _days_to_resolution(end_date: str) -> int:
-    """Return calendar days until market resolution, or 30 if unknown."""
+def _days_to_resolution(end_date: str, as_of: Optional[datetime] = None) -> int:
+    """Return calendar days until market resolution, or 30 if unknown.
+
+    as_of (Replay R1): reference clock for the computation.  None (production)
+    means wall-clock now; a replay run passes the archived cycle_ts so
+    days-to-resolution — and therefore the regime edge threshold — is computed
+    against the moment the decision was originally made.
+    """
     if not end_date:
         return 30  # conservative: treat as medium-term
     try:
         dt = datetime.fromisoformat(end_date.replace("Z", "+00:00"))
         if dt.tzinfo is None:
             dt = dt.replace(tzinfo=timezone.utc)
-        days = (dt - datetime.now(timezone.utc)).days
+        now = as_of if as_of is not None else datetime.now(timezone.utc)
+        days = (dt - now).days
         return max(0, days)
     except (ValueError, TypeError):
         return 30
@@ -457,10 +464,17 @@ class BayesianStrategy:
         market: Market,
         ext: ExternalSignals,
         occupied_families: set[str],
+        as_of: Optional[datetime] = None,
     ) -> Optional[TradingSignal]:
         """
         Evaluate a market and return a TradingSignal if actionable.
 
+        as_of (Replay R1): clock injection — None in production (wall-clock
+        now); a replay passes the archived cycle_ts so the regime threshold
+        matches the original decision moment.  Only days-to-resolution
+        depends on the clock; everything else is a pure function of
+        (market, ext, occupied_families) and the news/manifold clients.
+
         Returns None with a structured log line in all skip cases.
         Skip reasons (Phase 5 observability):
           SKIP_UNSUPPORTED  — category not supported
@@ -558,7 +572,7 @@ class BayesianStrategy:
             return None
 
         # ── Phase 4: regime min-edge ─────────────────────────────────────────
-        days = _days_to_resolution(market.end_date)
+        days = _days_to_resolution(market.end_date, as_of)
         regime_min = _regime_min_edge(category, days)
 
         # ── Bayesian probability estimation ──────────────────────────────────

docs/informe-final.md

@@ -0,0 +1,171 @@
+# Informe final de paper trading — polymarket-bot
+
+**Snapshot: 2026-07-06** · Generado desde la base de datos de producción con el bot
+aún operativo, como parte de la Fase 1 del plan de decomisión (bloqueo ISP de
+Polymarket en España, mayo 2026).
+
+---
+
+## ⚠️ Advertencia de tamaño muestral (léase primero)
+
+**12 trades ejecutados, de los cuales solo 2 llegaron a resolución, no constituyen
+evidencia estadística de nada.** Ni el P&L positivo demuestra que la estrategia
+funcione, ni su ausencia lo habría refutado. El propio bot lo sabía: sus criterios
+de promoción a dinero real exigían ≥10 mercados resueltos y ≥30 días de
+observación (`promotion_ready: false`, win rate y Sharpe deliberadamente
+reportados como `n/a (insufficient_sample)` en toda la instrumentación).
+
+El valor demostrable de este proyecto no está en el P&L: está en la
+infraestructura de evaluación, el pipeline de datos y la disciplina de gates que
+impidió que el bot tradease sin ventaja medible. Este informe presenta el P&L
+como lo que es — **una anécdota** — y dedica el grueso al sistema.
+
+---
+
+## 1. Resultados de trading (anecdóticos)
+
+Bankroll simulado: **$10.000** (USDC virtual) · Capital desplegado: **$2.447** ·
+Periodo de actividad: **14–22 de abril de 2026** (los 12 trades se abrieron en 9
+días; después, el endurecimiento de los gates de edge y confianza dejó el bot en
+0 trades — funcionando, evaluando ~1.350 ciclos/día, sin encontrar ventaja neta
+que superara sus propios umbrales).
+
+### P&L realizado: **+$247,78** (2 mercados resueltos)
+
+| Mercado | Dirección | Entrada | Resultado | P&L neto |
+|---|---|---|---|---|
+| Will Ken Paxton win the 2026 Texas Republican Primary? | BUY_YES | 0,618 | ✅ YES (2026-06-11) | **+$299,06** |
+| Will NVIDIA be the largest company in the world…? | BUY_NO | 0,156 | ❌ YES (2026-06-30) | **−$51,28** |
+
+P&L neto de comisiones (2% simulado por lado). 1 acierto / 1 fallo: n=2.
+
+### P&L no realizado (mark-to-market): **+$809,59** (5 posiciones abiertas)
+
+Marcado a los precios reales de Polymarket (Gamma API) el 2026-07-06. Estas
+posiciones **no están cerradas**: el número de abajo cambiaría cada día que los
+mercados sigan moviéndose, y varios no resuelven hasta noviembre de 2026.
+
+| Mercado | Dirección | Entrada | Precio actual (lado) | Coste neto | MTM P&L |
+|---|---|---|---|---|---|
+| Karen Bass gana la alcaldía de LA 2026 | BUY_YES | 0,268 | 0,595 | $510,00 | **+$600,07** |
+| OpenAI IPO antes del 31-dic-2026 | BUY_NO | 0,618 | 0,785 | $510,00 | **+$125,11** |
+| Demócratas ganan el Senado de Texas 2026 | BUY_YES | 0,418 | 0,435 | $509,49 | **+$10,32** |
+| Demócratas ganan gobernación de Ohio 2026 | BUY_NO | 0,458 | 0,520 | $407,74 | **+$46,12** |
+| Republicanos ganan gobernación de Nebraska 2026 | BUY_NO | 0,158 | 0,170 | $510,00 | **+$27,97** |
+
+> Nota metodológica: el tracker interno reportaba `unrealized_pnl_est = +$522,73`,
+> pero esa cifra es la **expectativa del propio modelo a fecha de entrada**
+> (edge_net × coste), no un marcado a mercado. La tabla de arriba usa precios
+> reales de mercado del día del snapshot, que es el criterio honesto. Ambas
+> cifras coinciden en el signo; ninguna de las dos es un resultado realizado.
+
+### Los otros 5 cierres (sin P&L): la historia real de abril
+
+De los 7 trades cerrados, solo 2 cerraron por resolución. Los otros 5 los cerró
+el propio bot al detectar defectos en su lógica — y son la parte más
+instructiva del histórico:
+
+- **3 por conflicto de familia**: dos posiciones sobre mercados hermanos del
+  mismo evento (p. ej. YES-demócratas y YES-republicanos en la misma
+  gobernación) — el agrupador de familias los detectó y cerró el lado débil.
+- **2 por el bug de inversión de Manifold**: la señal cruzada de Manifold se
+  aplicaba invertida en mercados espejo. Detectado por instrumentación, las
+  posiciones afectadas se cerraron y el matching se movió a modo
+  observacional-only con auditoría completa (165.538 filas en
+  `manifold_match_audit`).
+- **1 excluido de métricas** (`invalid_manifold_match_legacy`, P&L $0,00).
+
+Después de esa semana el bot no volvió a operar: los gates endurecidos
+(edge neto > mínimo por régimen, confianza ≥ 0,55, guardrail de noticias
+prior±0,25) filtraron el 100% de las ~223.000 evaluaciones posteriores.
+**Un sistema de paper trading cuyo resultado principal es "no tradees sin
+ventaja" es un resultado**, y está instrumentado hasta el último skip.
+
+---
+
+## 2. El sistema (el activo real)
+
+### Pipeline de señales
+
+| Métrica | Valor (snapshot 2026-07-06) |
+|---|---|
+| Ciclos de evaluación archivados | **5.356** (cadencia ~64 s) |
+| Evaluaciones mercado-ciclo archivadas | **222.738** |
+| Mercados distintos evaluados | 87 (events 114,9k · politics 96,0k · crypto/finance 7,7k · tech 4,2k evaluaciones) |
+| Archivo `signals` operativo desde | 2026-07-02 (replay R0) — crece ~55k filas/día |
+| Histórico de métricas diarias | **81 días** (2026-04-14 → 2026-07-06, `metrics_daily`) |
+| Snapshots de contexto externo | 5.356 (BTC, dominancia, Fear&Greed por ciclo) |
+
+Cada evaluación archiva el prior, la probabilidad estimada, el edge bruto/neto,
+la descomposición por feature (Fear&Greed, momentum, noticias, Manifold,
+dominancia BTC en log-odds), el gate que la bloqueó y el contexto de mercado —
+suficiente para reproducir la decisión completa offline.
+
+### Motor de replay (R0–R2, julio 2026)
+
+- Inyección de reloj + replay determinista de ciclos archivados:
+  **22.697/22.697 decisiones idénticas** al reproducir el histórico.
+- Joiner diario de resoluciones (CronJob 00:30 UTC) contra la Gamma API:
+  cobertura de outcomes 9/87 mercados al snapshot, creciendo sola.
+- Diseñado para calibrar sobre **todas** las evaluaciones (no solo trades),
+  esquivando el problema del n=12.
+
+### Estrategia (congelada desde 2026-07-03)
+
+Prior desde el precio de Polymarket → ajuste bayesiano en log-odds con señales
+externas (GNews con presupuesto de 5 consultas/día y guardrail prior±0,25;
+Fear&Greed; momentum BTC; Manifold observacional) → gates en cascada: prior
+extremo, edge bruto por régimen de volatilidad, edge neto tras comisión 2% y
+spread, confianza mínima 0,55, conflictos de familia, guard de reentrada.
+Sizing por Kelly fraccionado con techo por posición.
+
+### Infraestructura
+
+3 imágenes (bot / API FastAPI / dashboard) construidas por Gitea Actions solo
+cuando cambian sus fuentes, desplegadas por ArgoCD (prune + selfHeal) en k3s;
+PostgreSQL 16 en StatefulSet; secrets vía Infisical operator; smoke test
+PostSync contra la API; backups nocturnos con sync offsite (rclone → MEGA);
+observabilidad en Grafana + uptime-kuma + notificaciones Telegram (deploys,
+checkpoints del bot, fallos de sync). Retención de métricas y join de outcomes
+como CronJobs. **88 días de uptime del StatefulSet al snapshot.**
+
+Detalle honesto de observabilidad: el panel *Win Rate* del dashboard de Grafana
+(`docs/dashboard-grafana-2026-07-06.png`) pasó vacío los 81 días. No es una
+query rota: la API emite `win_rate: null` hasta tener ≥10 mercados resueltos
+(hubo 2), y Grafana no pinta nada con un null. El panel vacío es el criterio
+de promoción del bot aplicándose a su propia monitorización.
+
+---
+
+## 3. Estado de la base de datos al snapshot
+
+| Tabla | Filas | Contenido |
+|---|---|---|
+| `signals` | 222.549* | Archivo de evaluaciones por ciclo (desde 2026-07-02) |
+| `manifold_match_audit` | 165.538 | Auditoría completa del matching Polymarket↔Manifold |
+| `replay_decisions` | 22.697 | Decisiones reproducidas por el motor de replay |
+| `ext_snapshots` | 5.353* | Contexto externo por ciclo |
+| `metrics_daily` | 516 | Cierres diarios (81 días) + snapshots intradía de hoy |
+| `markets` | 87 | Universo evaluado |
+| `trades` | 12 | Ledger completo de paper trades |
+| `market_outcomes` | 9 | Resoluciones UMA-final joineadas |
+| `replay_runs` / `manifold_eval_cooldown` / `checkpoint_alerts` | 1 / 76 / 3 | Metadatos |
+
+\* La BD sigue viva; las tablas por ciclo crecen ~55k y ~1,3k filas/día
+respectivamente. Backup verificado del 2026-07-06 en
+`/data/backups/backups/polymarket-decommission/` (pg_dump -Fc + CSV.gz +
+checksums; restore de prueba con recuentos idénticos 11/11 tablas). El dump
+final se hará en la Fase 3, justo antes del apagado.
+
+---
+
+## 4. Conclusión
+
+El proyecto se archiva por una causa externa (bloqueo regulatorio del origen de
+datos), no por fracaso técnico. Lo que queda: un motor de evaluación bayesiano
+determinista y auditable, un pipeline de datos que archivó cada decisión con su
+contexto completo, y la evidencia — anecdótica en P&L, sólida en ingeniería —
+de que el sistema prefería no operar antes que operar sin ventaja. Todos los
+módulos centrales (edge, familias, replay, observabilidad) son agnósticos de la
+fuente y quedan listos para un eventual pivote a Manifold
+(`docs/pivot-manifold.md`, Fase 5).

requirements.txt

@@ -2,7 +2,7 @@
 asyncpg==0.29.0
 httpx==0.27.0
 fastapi==0.111.0
-uvicorn[standard]==0.29.0
+uvicorn[standard]==0.50.2
 pydantic==2.7.0
 
 # Polymarket (install from PyPI when ready for real trading)

tests/test_outcomes.py

@@ -0,0 +1,174 @@
+"""Replay R2 tests — outcome fetching and calibration scoring."""
+import asyncio
+import math
+
+import pytest
+
+from bot.data.polymarket import MarketResolution
+from bot.outcomes import (
+    LOGLOSS_EPS,
+    compute_calibration,
+    fetch_outcomes,
+    print_report,
+)
+from datetime import datetime, timezone
+
+
+class FakePoly:
+    """get_market_resolution stand-in driven by a dict of canned responses."""
+
+    def __init__(self, responses: dict):
+        self.responses = responses
+        self.calls: list[str] = []
+
+    async def get_market_resolution(self, market_id: str):
+        self.calls.append(market_id)
+        return self.responses.get(market_id)
+
+
+RESOLVED_AT = datetime(2026, 7, 1, 12, 0, tzinfo=timezone.utc)
+
+
+def _row(market_id="m1", category="politics", est=0.6, prior=0.5, outcome=1.0):
+    return {
+        "market_id": market_id,
+        "category": category,
+        "estimated_prob": est,
+        "prior_prob": prior,
+        "outcome": outcome,
+    }
+
+
+# ── fetch_outcomes ───────────────────────────────────────────────────────────
+
+def test_fetch_keeps_only_definitive_resolutions():
+    poly = FakePoly({
+        "yes":      MarketResolution(resolved=True, resolution=1.0,
+                                     resolved_at=RESOLVED_AT),
+        "no":       MarketResolution(resolved=True, resolution=0.0,
+                                     resolved_at=None),
+        "open":     MarketResolution(resolved=False),
+        "disputed": MarketResolution(resolved=False),
+        "apierror": None,  # get_market_resolution returns None on HTTP errors
+    })
+    out = asyncio.run(
+        fetch_outcomes(poly, ["yes", "no", "open", "disputed", "apierror"])
+    )
+    assert poly.calls == ["yes", "no", "open", "disputed", "apierror"]
+    assert out == [
+        {"market_id": "yes", "outcome": 1.0, "resolved_at": RESOLVED_AT},
+        {"market_id": "no", "outcome": 0.0, "resolved_at": None},
+    ]
+
+
+def test_fetch_empty_list_is_noop():
+    poly = FakePoly({})
+    assert asyncio.run(fetch_outcomes(poly, [])) == []
+    assert poly.calls == []
+
+
+# ── compute_calibration ──────────────────────────────────────────────────────
+
+def test_no_rows_returns_none():
+    assert compute_calibration([]) is None
+
+
+def test_single_row_known_values():
+    m = compute_calibration([_row(est=0.8, prior=0.6, outcome=1.0)])
+    assert m["n_evaluations"] == 1
+    assert m["n_markets"] == 1
+    assert m["brier_model"] == pytest.approx((0.8 - 1.0) ** 2)
+    assert m["brier_prior"] == pytest.approx((0.6 - 1.0) ** 2)
+    assert m["logloss_model"] == pytest.approx(-math.log(0.8))
+    assert m["logloss_prior"] == pytest.approx(-math.log(0.6))
+    # one market: macro == micro
+    assert m["brier_model_macro"] == pytest.approx(m["brier_model"])
+    assert m["brier_prior_macro"] == pytest.approx(m["brier_prior"])
+
+
+def test_logloss_no_outcome_branch():
+    m = compute_calibration([_row(est=0.2, prior=0.7, outcome=0.0)])
+    assert m["logloss_model"] == pytest.approx(-math.log(0.8))
+    assert m["logloss_prior"] == pytest.approx(-math.log(0.3))
+
+
+def test_logloss_clipping_keeps_hard_miss_finite():
+    # A hard 1.0 estimate on a NO outcome must not produce inf.
+    m = compute_calibration([_row(est=1.0, prior=0.5, outcome=0.0)])
+    assert math.isfinite(m["logloss_model"])
+    assert m["logloss_model"] == pytest.approx(-math.log(LOGLOSS_EPS))
+
+
+def test_micro_weights_evaluations_macro_weights_markets():
+    # Market a: 3 evaluations, model error 0.1; market b: 1 evaluation, error 0.5.
+    rows = [
+        _row(market_id="a", est=0.9, prior=0.8, outcome=1.0),
+        _row(market_id="a", est=0.9, prior=0.8, outcome=1.0),
+        _row(market_id="a", est=0.9, prior=0.8, outcome=1.0),
+        _row(market_id="b", est=0.5, prior=0.6, outcome=1.0),
+    ]
+    m = compute_calibration(rows)
+    assert m["n_evaluations"] == 4
+    assert m["n_markets"] == 2
+    # micro: (3*0.01 + 0.25) / 4 ; macro: (0.01 + 0.25) / 2
+    assert m["brier_model"] == pytest.approx((3 * 0.01 + 0.25) / 4)
+    assert m["brier_model_macro"] == pytest.approx((0.01 + 0.25) / 2)
+    assert m["brier_prior"] == pytest.approx((3 * 0.04 + 0.16) / 4)
+    assert m["brier_prior_macro"] == pytest.approx((0.04 + 0.16) / 2)
+
+
+def test_model_beating_market_gives_negative_delta():
+    # est closer to the outcome than the price on every row
+    rows = [
+        _row(market_id="a", est=0.8, prior=0.6, outcome=1.0),
+        _row(market_id="b", est=0.3, prior=0.45, outcome=0.0),
+    ]
+    m = compute_calibration(rows)
+    assert m["brier_model"] < m["brier_prior"]
+    assert m["logloss_model"] < m["logloss_prior"]
+
+
+def test_per_category_grouping_and_unknown():
+    rows = [
+        _row(market_id="a", category="politics", est=0.8, prior=0.6, outcome=1.0),
+        _row(market_id="b", category="politics", est=0.7, prior=0.6, outcome=1.0),
+        _row(market_id="c", category=None, est=0.4, prior=0.5, outcome=0.0),
+    ]
+    m = compute_calibration(rows)
+    assert set(m["per_category"]) == {"politics", "unknown"}
+    pol = m["per_category"]["politics"]
+    assert pol["n"] == 2 and pol["markets"] == 2
+    assert pol["brier_model"] == pytest.approx((0.04 + 0.09) / 2)
+    unk = m["per_category"]["unknown"]
+    assert unk["n"] == 1 and unk["markets"] == 1
+    assert unk["brier_model"] == pytest.approx(0.16)
+
+
+def test_repeated_market_counts_once_in_markets():
+    rows = [
+        _row(market_id="a", est=0.8, prior=0.6, outcome=1.0),
+        _row(market_id="a", est=0.7, prior=0.55, outcome=1.0),
+    ]
+    m = compute_calibration(rows)
+    assert m["n_markets"] == 1
+    assert m["per_category"]["politics"]["markets"] == 1
+
+
+# ── print_report ─────────────────────────────────────────────────────────────
+
+def test_report_handles_no_metrics(capsys):
+    print_report(None, "R0 archive")
+    assert "no scorable rows yet" in capsys.readouterr().out
+
+
+def test_report_prints_all_metric_lines(capsys):
+    m = compute_calibration([
+        _row(market_id="a", est=0.8, prior=0.6, outcome=1.0),
+        _row(market_id="b", category=None, est=0.4, prior=0.5, outcome=0.0),
+    ])
+    print_report(m, "R0 archive")
+    out = capsys.readouterr().out
+    assert "2 evaluations, 2 markets" in out
+    for label in ("Brier micro", "Brier macro", "logloss micro",
+                  "politics", "unknown"):
+        assert label in out

tests/test_replay_engine.py

@@ -0,0 +1,367 @@
+"""
+Tests for the Replay R1 replay core (bot/replay.py) and the as_of clock
+injection in BayesianStrategy.evaluate().
+
+The central contract is round-trip fidelity: a decision recorded by R0 and
+replayed through replay_cycle() with the same strategy constants must match
+field-for-field (matched=True, mismatch_field=None).  Each round-trip test
+produces the "archive" by running the real evaluate() with FakeNews, then
+replays the drained record as if it had been read back from the signals table.
+"""
+import asyncio
+from datetime import datetime, timedelta, timezone
+
+import pytest
+
+import bot.strategy.bayesian as bayesian
+from bot.data.polymarket import Market, market_family_key
+from bot.strategy.bayesian import BayesianStrategy, _days_to_resolution
+from bot.replay import (
+    ReplayNews,
+    build_ext,
+    build_market,
+    replay_cycle,
+    strategy_config_hash,
+)
+
+from tests.test_news_guardrail import FakeNews, _sentiment_for
+
+
+def _end_date(days_ahead: int = 20) -> str:
+    dt = datetime.now(timezone.utc) + timedelta(days=days_ahead)
+    return dt.strftime("%Y-%m-%dT00:00:00Z")
+
+
+def _make_market(
+    yes_price: float,
+    question: str = "Will John Smith win the election?",
+    category: str = "politics",
+    market_id: str = "mkt-replay-1",
+    end_date: str = None,
+) -> Market:
+    return Market(
+        id=market_id,
+        condition_id="cond-replay-1",
+        question=question,
+        yes_token_id="yes-tok",
+        no_token_id="no-tok",
+        yes_price=yes_price,
+        no_price=1.0 - yes_price,
+        volume_24h=50_000.0,
+        end_date=end_date if end_date is not None else _end_date(),
+        active=True,
+        category=category,
+    )
+
+
+def _snapshot(valid: bool = True) -> dict:
+    """An ext_snapshots row as read back from the DB."""
+    return {
+        "btc_price": 100_000.0,
+        "btc_change_24h": 0.0,
+        "eth_price": 4_000.0,
+        "eth_change_24h": 0.0,
+        "btc_dominance": 50.0,
+        "fear_greed_index": 50,
+        "fear_greed_label": "neutral",
+        "total_market_cap_change": 0.0,
+        "valid": valid,
+    }
+
+
+def _market_row(market: Market) -> dict:
+    """A markets-table row for the given Market."""
+    return {
+        "id": market.id,
+        "condition_id": market.condition_id,
+        "question": market.question,
+        "category": market.category,
+        "end_date": market.end_date,
+    }
+
+
+def _record_with_live_evaluate(
+    market: Market,
+    news=None,
+    families: set = frozenset(),
+) -> dict:
+    """Run the real evaluate() and return the R0 record it produced —
+    the same dict save_signal_records() would have archived."""
+    strategy = BayesianStrategy(news=news, manifold=None, db=None)
+    asyncio.run(strategy.evaluate(market, build_ext(_snapshot()), set(families)))
+    return strategy.drain_cycle_records()[0]
+
+
+def _replay_one(record: dict, market: Market, snapshot: dict = None) -> dict:
+    cycle_ts = datetime.now(timezone.utc)
+    decisions = asyncio.run(replay_cycle(
+        cycle_ts,
+        snapshot or _snapshot(),
+        [record],
+        {market.id: _market_row(market)},
+    ))
+    assert len(decisions) == 1
+    return decisions[0]
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Clock injection
+# ─────────────────────────────────────────────────────────────────────────────
+
+def test_days_to_resolution_uses_injected_clock():
+    end = "2026-08-01T00:00:00Z"
+    as_of = datetime(2026, 7, 2, 12, 0, tzinfo=timezone.utc)
+    assert _days_to_resolution(end, as_of) == 29
+    assert _days_to_resolution(end, as_of - timedelta(days=60)) == 89
+
+
+def test_default_clock_is_wall_clock():
+    end = _end_date(days_ahead=40)
+    assert _days_to_resolution(end) == _days_to_resolution(
+        end, datetime.now(timezone.utc)
+    )
+
+
+def test_as_of_changes_regime_threshold():
+    """Same politics market: <30 d out → regime 0.08; replayed from 60 d
+    earlier → regime 0.12.  The clock, not the wall time, must decide."""
+    market = _make_market(0.470)
+    sentiment = _sentiment_for(0.470, 0.601)
+
+    def _regime(as_of):
+        strategy = BayesianStrategy(news=FakeNews(sentiment), manifold=None, db=None)
+        asyncio.run(strategy.evaluate(
+            market, build_ext(_snapshot()), set(), as_of=as_of,
+        ))
+        return strategy.drain_cycle_records()[0]["regime_min_edge"]
+
+    now = datetime.now(timezone.utc)
+    assert _regime(now) == pytest.approx(0.08)
+    assert _regime(now - timedelta(days=60)) == pytest.approx(0.12)
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Round-trip fidelity: record with live evaluate(), replay, expect match
+# ─────────────────────────────────────────────────────────────────────────────
+
+def test_roundtrip_confidence_skip():
+    """Georgia signature: edge passes, confidence blocks — full-field match."""
+    sentiment = _sentiment_for(0.470, 0.601)
+    market = _make_market(0.470)
+    record = _record_with_live_evaluate(market, news=FakeNews(sentiment))
+    assert record["skip_reason"] == "confidence"
+
+    decision = _replay_one(record, market)
+    assert decision["matched"] is True
+    assert decision["mismatch_field"] is None
+    assert decision["skip_reason"] == "confidence"
+    assert decision["estimated_prob"] == pytest.approx(record["estimated_prob"])
+    assert decision["edge_net"] == pytest.approx(record["edge_net"])
+    assert decision["confidence"] == pytest.approx(record["confidence"])
+    assert decision["direction"] == record["direction"]
+    assert decision["would_trade"] is False
+
+
+def test_roundtrip_edge_net_skip():
+    market = _make_market(0.50)
+    record = _record_with_live_evaluate(market)
+    assert record["skip_reason"] == "edge_net"
+
+    decision = _replay_one(record, market)
+    assert decision["matched"] is True
+    assert decision["would_trade"] is False
+
+
+def test_roundtrip_guardrail_clamp():
+    """Clamped posterior must reproduce exactly (raw != final in archive)."""
+    market = _make_market(0.845)
+    record = _record_with_live_evaluate(
+        market, news=FakeNews(_sentiment_for(0.845, 0.431))
+    )
+    assert record["guardrail_applied"] is True
+
+    decision = _replay_one(record, market)
+    assert decision["matched"] is True
+    assert decision["raw_final_prob"] == pytest.approx(record["raw_final_prob"])
+    assert decision["estimated_prob"] == pytest.approx(record["estimated_prob"])
+
+
+def test_roundtrip_prior_extreme():
+    market = _make_market(0.03)
+    record = _record_with_live_evaluate(market)
+    assert record["skip_reason"] == "prior_extreme"
+
+    decision = _replay_one(record, market)
+    assert decision["matched"] is True
+    assert decision["skip_reason"] == "prior_extreme"
+
+
+def test_roundtrip_family_skip():
+    """Family-skipped rows replay with their own family injected as occupied."""
+    market = _make_market(0.50)
+    record = _record_with_live_evaluate(
+        market, families={market_family_key(market)}
+    )
+    assert record["skip_reason"] == "family"
+
+    decision = _replay_one(record, market)
+    assert decision["matched"] is True
+    assert decision["skip_reason"] == "family"
+
+
+def test_roundtrip_unsupported():
+    market = _make_market(0.50, question="Will it rain tomorrow?", category="")
+    record = _record_with_live_evaluate(market)
+    assert record["skip_reason"] == "unsupported"
+
+    decision = _replay_one(record, market)
+    assert decision["matched"] is True
+
+
+def test_roundtrip_no_signals():
+    """ext.valid=False archived → replay rebuilds the invalid snapshot."""
+    market = _make_market(0.50)
+    strategy = BayesianStrategy(news=None, manifold=None, db=None)
+    asyncio.run(strategy.evaluate(market, build_ext(_snapshot(valid=False)), set()))
+    record = strategy.drain_cycle_records()[0]
+    assert record["skip_reason"] == "no_signals"
+
+    decision = _replay_one(record, market, snapshot=_snapshot(valid=False))
+    assert decision["matched"] is True
+
+
+def test_roundtrip_trade_path(monkeypatch):
+    """skip_reason=None (tradeable) round-trips with would_trade=True.
+    Politics can't clear MIN_CONFIDENCE=0.55 (the known ceiling), so the
+    gate is lowered for this test only — both record and replay see the
+    same constant, which is exactly the config_hash contract."""
+    monkeypatch.setattr(bayesian, "MIN_CONFIDENCE", 0.45)
+    sentiment = _sentiment_for(0.470, 0.601)
+    market = _make_market(0.470)
+    record = _record_with_live_evaluate(market, news=FakeNews(sentiment))
+    assert record["skip_reason"] is None
+
+    decision = _replay_one(record, market)
+    assert decision["matched"] is True
+    assert decision["skip_reason"] is None
+    assert decision["would_trade"] is True
+    assert decision["direction"] == "BUY_YES"
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Replay-specific semantics
+# ─────────────────────────────────────────────────────────────────────────────
+
+def test_budget_skipped_row_replays_without_news():
+    """A budget-skipped archive row (sentiment 0.0) must replay to the same
+    no-news decision — and never consume a replay-side budget."""
+    market = _make_market(0.50)
+    strategy = BayesianStrategy(news=FakeNews(0.9), manifold=None, db=None)
+    strategy._news_queries_this_cycle = bayesian.MAX_NEWS_QUERIES_PER_CYCLE
+    asyncio.run(strategy.evaluate(market, build_ext(_snapshot()), set()))
+    record = strategy.drain_cycle_records()[0]
+    assert record["news_budget_skipped"] is True
+    assert record["news_sentiment"] == 0.0
+
+    decision = _replay_one(record, market)
+    assert decision["matched"] is True
+    assert decision["estimated_prob"] == pytest.approx(record["estimated_prob"])
+
+
+def test_reentry_guard_row_is_recalibrated_not_compared():
+    """record_skip() rows carry no decision fields; the replay re-evaluates
+    them (calibration data) but marks them non-comparable."""
+    market = _make_market(0.50)
+    strategy = BayesianStrategy(news=None, manifold=None, db=None)
+    strategy.record_skip(market, "reentry_guard")
+    record = strategy.drain_cycle_records()[0]
+
+    decision = _replay_one(record, market)
+    assert decision["matched"] is None
+    assert decision["recorded_skip_reason"] == "reentry_guard"
+    # Re-evaluated on its merits: a full decision despite the recorded skip
+    assert decision["estimated_prob"] is not None
+    assert decision["skip_reason"] == "edge_net"
+
+
+def test_missing_market_row_flagged_not_crashed():
+    market = _make_market(0.50)
+    record = _record_with_live_evaluate(market)
+
+    decisions = asyncio.run(replay_cycle(
+        datetime.now(timezone.utc), _snapshot(), [record], {},
+    ))
+    assert decisions[0]["matched"] is False
+    assert decisions[0]["mismatch_field"] == "market_missing"
+
+
+def test_mismatch_detected_when_config_differs(monkeypatch):
+    """Counterfactual sanity: replaying under a different guardrail band
+    must produce matched=False with the diverging field named."""
+    market = _make_market(0.845)
+    record = _record_with_live_evaluate(
+        market, news=FakeNews(_sentiment_for(0.845, 0.431))
+    )
+    assert record["guardrail_applied"] is True
+
+    monkeypatch.setattr(bayesian, "MAX_NEWS_ONLY_PROB_SHIFT", 0.10)
+    decision = _replay_one(record, market)
+    assert decision["matched"] is False
+    # Tighter clamp (prior 0.845 ± 0.10 → est 0.745): edge_net drops from
+    # 0.21 to 0.06 < regime 0.08, so the skip flips confidence → edge_net
+    # and skip_reason is the first field _compare() sees diverge.
+    assert decision["mismatch_field"] == "skip_reason"
+    assert decision["skip_reason"] == "edge_net"
+
+
+def test_multi_row_cycle_preserves_order_and_isolation():
+    """Rows replay independently within a cycle: a family skip and a full
+    evaluation with different sentiments don't bleed into each other."""
+    m1 = _make_market(0.470, market_id="m1")
+    m2 = _make_market(
+        0.50, market_id="m2",
+        question="Will Jane Doe win the Georgia Senate race?",
+    )
+    r1 = _record_with_live_evaluate(m1, news=FakeNews(_sentiment_for(0.470, 0.601)))
+    r2 = _record_with_live_evaluate(m2)  # no news → edge_net skip
+
+    decisions = asyncio.run(replay_cycle(
+        datetime.now(timezone.utc),
+        _snapshot(),
+        [r1, r2],
+        {"m1": _market_row(m1), "m2": _market_row(m2)},
+    ))
+    assert [d["market_id"] for d in decisions] == ["m1", "m2"]
+    assert all(d["matched"] is True for d in decisions)
+    assert decisions[0]["skip_reason"] == "confidence"
+    assert decisions[1]["skip_reason"] == "edge_net"
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Run tagging
+# ─────────────────────────────────────────────────────────────────────────────
+
+def test_config_hash_stable_and_sensitive(monkeypatch):
+    h1 = strategy_config_hash()
+    assert strategy_config_hash() == h1
+    monkeypatch.setattr(bayesian, "MAX_NEWS_ONLY_PROB_SHIFT", 0.10)
+    assert strategy_config_hash() != h1
+
+
+def test_replay_news_returns_current_sentiment():
+    news = ReplayNews()
+    assert asyncio.run(news.get_sentiment("q")) == 0.0
+    news.sentiment = -0.42
+    assert asyncio.run(news.get_sentiment("q")) == -0.42
+
+
+def test_build_market_reconstruction():
+    market = _make_market(0.37)
+    record = _record_with_live_evaluate(market)
+    rebuilt = build_market(_market_row(market), record)
+    assert rebuilt.id == market.id
+    assert rebuilt.yes_price == pytest.approx(0.37)
+    assert rebuilt.volume_24h == pytest.approx(market.volume_24h)
+    assert rebuilt.end_date == market.end_date
+    assert rebuilt.category == "politics"
+    assert market_family_key(rebuilt) == market_family_key(market)