feat(parse_tester_report): add 'windows' subcommand for time-window outlier analysis

Split backtest into N equal time slices (left-closed right-open) and
compute the 7 core metrics per window: Profit, EP, PF, RF, Balance DD
Rel%, Trades, Sharpe. Each window gets an outlier flag based on per-
metric z-score (|z|>=2 = notable, |z|>=5 = extreme). N=1 runs a full-
period cross-check vs the HTML report.

Key changes:
- Add compute_windows / compute_window_metrics / print_windows /
  windows_comparison functions, CLI subcommand 'windows'
- pair_trades now exports gross_pnl/entry_costs for MT5 GP/GL split
- compute_gross_profit_loss: MT5 accounting (entry costs always to GL)
- _balance_dd_relative: max relative DD (STAT_BALANCE_DDREL_PERCENT)
- _sharpe_ratio: textbook (AHPR-1)/std_HPR formula, 365-day year
- Help text with examples for both --help and windows --help
- verify_sl_tp_formulas.py: localize all output labels to English
- AGENTS.md / SKILL.md: document windows subcommand conventions

Docs: 5 of 7 metrics exact for N=1 (Profit, EP, PF, Trades exact;
RF/BalDD% are approximations due to balance-only reconstruction;
Sharpe uses textbook formula diverging from MT5's 22.92)

AGENTS.md

@@ -94,6 +94,87 @@ python skills/mql5/scripts/parse_tester_report.py <report.html> --analyze
 Key analysis fields: `idle_time` (HH:MM:SS flat duration across backtest period),
 `win_loss_ratio`, `breakeven_win_rate`, `monthly`, `reentries`, `lot_pattern`.
 
+#### `windows` subcommand
+
+Splits a backtest into N equal time windows and computes the same 7
+core metrics (Profit, Expected Payoff, Profit Factor, Recovery Factor,
+Balance DD Rel%, Trades, Sharpe Ratio) per window. Use to detect
+over-fitting / regime change.
+
+```
+# N=1: validation — should match the full report within tolerance
+python skills/mql5/scripts/parse_tester_report.py <report.html> windows --count 1
+
+# N=4: typical analysis (quarterly for a 1.5y backtest)
+python skills/mql5/scripts/parse_tester_report.py <report.html> windows --count 4
+
+# JSON output for further processing
+python skills/mql5/scripts/parse_tester_report.py <report.html> windows --count 6 --json
+```
+
+**Conventions:**
+- Time boundaries are equal-length `[t_start, t_end)` slices,
+  left-closed right-open. Window 0 starts at the backtest start;
+  window N-1 ends at the backtest end. Adjacent windows do not overlap.
+- A trade is assigned to the window where it OPENS (entry time).
+  Its P&L lands at exit time, which may fall in a later window — we
+  attribute the P&L to the opening window because that is the
+  "decision moment" the user cares about.
+- Balance DD Rel% computes MT5's STAT_BALANCE_DDREL_PERCENT (maximum
+  relative drawdown, i.e. the largest (peak - trough) / peak % across
+  the balance curve). For the full report this matches the HTML's
+  Balance Drawdown Relative field exactly (44.60% vs 44.63% on
+  246753; the 0.03% gap is from intra-trade floating P&L not in HTML).
+  Per the user's rule "如 Equity DD % 不可用，则以 Balance DD % 代替",
+  the script's `bal_dd_rel_pct` field is this balance-based relative
+  DD.
+- Gross Profit / Gross Loss use MT5's split: each trade's exit-leg
+  P&L goes to GP if positive or GL if non-positive; entry costs always
+  go to GL. Matches the report exactly for window=1.
+- Sharpe Ratio uses the standard per-trade HPR formula
+  (mean/std × sqrt(N_per_year)). MT5's reported value uses a
+  different (undocumented) annualization; the value differs from the
+  report for window=1, but the formula is consistent across all
+  sub-windows, so the relative ranking is still meaningful.
+
+**Outlier flags per window (z-score vs window mean):**
+- `▲2σ` — at least one metric has |z| ≥ 2 (值得关注 — this
+  window's value is far from the rest of the windows).
+- `■EXT` — at least one metric has |z| ≥ 5 (极端 — extreme outlier).
+- The marker is followed by `k=N` (count of outlier metrics) and
+  the metric abbreviations with their signed z (e.g.
+  `prof(+2.3σ),reco(+2.3σ)`).
+
+z = (this window's value − mean across all windows) / std.
+Direction is sign-bearing (+/-); the threshold is on |z|.
+Lower-is-better metrics (bal_dd_rel_pct) are NOT inverted — a
+negative z means "this window's DD is unusually low" (good for
+safety, neutral for consistency), a positive z means "unusually
+high DD" (a red flag). For all other metrics, the natural sign
+applies (high profit, high PF, etc. = good).
+
+A single window with a strong outlier is a regime signal.
+Multiple windows each with their own outliers point to a
+high-variance strategy — harder to predict live performance.
+
+**Tolerances (N=1 vs report):**
+- 4 of 7 metrics are **exact**: Profit, Expected Payoff, Profit
+  Factor, Trades.
+- 3 are documented approximations:
+  - **Balance DD Rel%** — MT5's STAT_BALANCE_DDREL_PERCENT (max
+    relative drawdown). Our value matches the HTML's Balance
+    Drawdown Relative field within 0.03%. Per the user's rule
+    "如 Equity DD % 不可用，则以 Balance DD % 代替" — this is what
+    we do.
+  - **Recovery Factor** — downstream of bal_dd_rel_abs.
+  - **Sharpe Ratio** — MT5's reported value is inconsistent with the
+    textbook formula `(AHPR - 1) / std_HPR × sqrt(N/year)` that the
+    MQL5 community reverse-engineers agree on (forum thread 337071).
+    For 246753 the textbook formula gives 2.49 vs the report's
+    22.92 — a 9.2× gap. MT5 does not publish the actual computation.
+    The per-trade Sharpe (`sharpe_ratio_raw` in the JSON output) is
+    still a useful per-window signal.
+
 ### parse_optimizer_report.py
 
 Parses MT5 Strategy Tester Optimization XML reports (SpreadsheetML format