fix(engine): freeze default percent_of_equity/cash market-order sizing at the signal bar

[luisleo526]

Summary

TradingView sizes a DEFAULT-qty (qty=na) market order at the signal bar, not at the fill:

equity_S     = initial_capital + realized net profit + open_profit(close(S))
sizing_price = close(S) + slippage*mintick*(+1 buy / -1 sell)
qty          = floor_step(reserve_percent_commission(equity_S * pct/100) / fx / (sizing_price * pointvalue))

The order then fills at the next bar's open carrying that frozen quantity. Evaluating the same expression at fill time — the previous behaviour — was wrong three separable ways on a reversal or a gapped bar:

1. Double count — flip_market_position_to() books emit_close_trade() (adding the closing lot's PnL to net_profit_sum_) and only then calls calc_qty_for_type(), while position_* still holds the lot just closed, so open_profit() re-adds that PnL as a phantom mark.
2. Look-ahead — open_profit() marked at current_bar_.close, i.e. the fill bar's close, a price unknown when the order fills at that bar's open.
3. Wrong divisor — divided by fill_price = open(S+1) rather than close(S). 49.9% of 15m bars in the reference window have open != prev close.

Freezing at placement removes all three at once: current_bar_ is the signal bar there.

Evidence

A flat entry isolates the divisor (no position, so the double count cannot apply). Corpus golden analyzer-parity-percent-of-equity-sizing-01 entry #1: close(S)=1807.83, open(S+1)=1807.82, capital 1e6, pct 99 → TV Size(qty) = 547.6178 = 990000/1807.83. Base engine emitted 547.6209 = 990000/1807.82. Patched engine emits 547.6180.

The buggy fill-time formula predicts 1739/1739 reversal quantities in one exercised strategy; close(S) predicts 370/370 gapped reversals and 507/507 gapped flat entries where open(S+1) predicts 0.

Why a new field

The frozen quantity is stored in a new PendingOrder::frozen_default_qty, not in qty. qty doubles as the "an explicit qty was provided" flag, and several sites recover "was this default-sized?" from std::isnan(o.qty): reduce_oca_group's default-sized cancel, the pending-reversal-entry binding, the OCA fully-filled heuristic, and the partial-vs-full exit classification. An earlier revision wrote into qty and silently flipped all of them — it disabled reversal binding for exactly the orders this change targets, and cost 5 downward tier moves. qty now stays NaN; the frozen value is read only where a quantity is actually computed.

Scope

MARKET (strategy.entry) and RAW market (strategy.order) placements with default PERCENT_OF_EQUITY / CASH sizing. FIXED default sizing needs no freeze (identical value at fill). Priced (limit/stop) entries keep the legacy fill-time sizing — TV's basis for an order armed bars before its fill is not established.

Known residual (deliberately not addressed)

For an entry sized while a position is open with commission_type=percent, TV may additionally deduct the entry fee already charged on the open lot. Independent reconstructions of TV's equity chain disagree, and the end-to-end evidence is confounded by an unrelated order-admission gap. An in-code comment warns against re-adding it by algebra; it needs a clean-room TV probe.

Test plan

• ctest — 80/80 (adds tests/test_default_qty_signal_freeze.cpp, pinning the reversal, gap-down and gap-up flat entries, process_orders_on_close, and CASH; a gap must neither re-size nor decline a flat, fully-affordable entry)
• scripts/run_corpus.sh — 252/252 ran, excellent=239 strong=12 anomaly=1, the unchanged baseline
• Downstream validation sweep of 409 community strategies — 5 tier upgrades, 0 downgrades

🤖 Generated with Claude Code