Trading metrics + equity curve in backtest report (ABI v2)

[luisleo526]

What

Adds a comprehensive trading-metrics suite and the raw per-script-bar equity curve to pf_report_t, validated against a real TradingView export and two independent quant libraries.

• New ABI surface (PF_ABI_VERSION 2): pf_metrics_t (trade-stat blocks for All/Long/Short + equity stats) and pf_equity_point_t* equity_curve appended to pf_report_t; commission / entry_bar_index / exit_bar_index appended to pf_trade_t. New pf_abi_version() export — pf_report_t is caller-allocated, so every harness now asserts the version before running (old mirror = silent stack corruption).
• Metrics: net/gross profit ±%, profit factor, win rate, avg/largest win/loss ±%, expectancy, win/loss streaks, avg bars in trade, commission paid; max equity drawdown/run-up ±%, buy & hold, TV-method monthly Sharpe/Sortino (RF 2%, chart-tz month bucketing) + per-bar density-annualized variants, CAGR, Calmar, recovery factor, time in market, open P&L. Conventions pinned per-field in pineforge.h doxygen.
• Collection: one equity point per script bar with magnifier-invariant timestamps (ab.bar.timestamp; pinned by a magnifier on/off bit-identity test), reset-safe for handle reuse.
• Computation: pure functions in src/engine_metrics.cpp (no engine dependency), dd/run-up walk mirrors update_equity_extremes (enforced by integration test + reciprocal comments).
• TV-arbitrated corrections (vs a real Strategy Tester export of composite-4emarsi-integration-01, 336 trades): per-trade pnl_pct is now net return-on-cost (closes long-standing O5 finding; TV trade #258: 102.44 USD / 2276.66 entry ⇒ 4.50%, old price-ratio formula said 4.72%); largest_*_pct tracked independently of the largest-USD trade; avg_bars counted inclusively.
• All Python ctypes mirrors updated + ABI-guarded: scripts/run_strategy.py, tutorials, docker/run_json.py, bench script. Docs FFI pages (Python/Rust) and report schema synced.
• New optional scripts/crossvalidate_metrics.py (venv-based, no repo deps).

Validation

• ctest: 71/71 (new test_metrics: 1200+ checks incl. closed-form Sharpe/Sortino oracles 19/20 and 114/61, non-UTC month bucketing, magnifier bit-identity, handle-reuse reset, engine-vs-walk drawdown invariant). ASan clean.
• Corpus: full sweep holds the exact baseline — excellent=245, anomaly=1; regenerated trade CSVs differ only in the Net PnL % column (expected from the pnl_pct correction; verified across all 158 changed files).
• Bench: with-magnifier hot loop at parity vs same-machine baseline control (median ratio 0.995); collection is one push_back per script bar outside the sub-bar loop.
• TradingView: every field both sides compute in the exported panels matches within TV 2-dp rounding (counts/PF/percent-profitable exact; USD deltas ≤0.34 total from known 2-trade price drift).
• quantstats 0.0.81 + empyrical-reloaded 0.5.12: max DD ±%, Sharpe/Sortino (both variants), CAGR, Calmar, recovery factor all match to ≤1e-11 relative on two strategies; every residual delta proven to be a library convention (geometric vs arithmetic RF, abs()-based recovery).

Known deltas / follow-ups

• TV "Max run-up (close-to-close)" 426.44 vs engine 406.33 — run-up definition difference (engine mirrors its pre-existing trough-reset-on-new-peak semantics); needs TV Run-ups/Drawdowns tab semantics pinned before changing anything.
• TV Sharpe/Sortino and max-DD panel values still unexported (Risk-adjusted performance / Drawdowns tabs) — engine values are library-validated meanwhile.
• Sibling repo pineforge-codegen-oss branch feat/abi-v2-metrics (compile-test fix for the new version.h include) must merge in lockstep; codegen-mcp docker image needs a rebuild.
• Corpus submodule has regenerated engine_trades.csv files (Net PnL % column) — submodule bump intentionally left out of this PR.

🤖 Generated with Claude Code