bug(fills): double-slippage on close-then-enter (close_only_opposite) path #27

Summary

close_opposite_then_enter applies slippage twice to the close leg when a flat-issued opposite-direction entry closes an existing position. Latent: masked by the corpus (all probes run slippage=0, where directional mintick rounding is idempotent).

Root cause

src/engine_orders.cpp:

// close_opposite_then_enter (~:539)
double exit_fill = apply_slippage(fill_price, position_side_ == PositionSide::SHORT); // slip #1
execute_partial_exit_qty(exit_fill, close_qty);

But execute_partial_exit_qty slips again internally (~:155):

void BacktestEngine::execute_partial_exit_qty(double fill_price, double qty_to_close) {
    ...
    fill_price = apply_slippage(fill_price, is_buy);   // slip #2

The helper's contract is "receives an un-slipped fill_price and slips internally" — confirmed by its other caller execute_partial_exit_by_percent (~:168) which passes the raw fill_price. close_opposite_then_enter pre-slips, so the close fills at price ± 2·slippage·mintick.

Trigger (the close_only_opposite path)

Set in apply_entry_order_fill (engine_fills.cpp): created_position_side == FLAT && position_side_ != FLAT && entry_req != position_side_ — i.e. a flat-issued (bracket-style) entry in the opposite direction to the open position fires and closes it (probe 80–87 family). Only bites when slippage > 0.

Impact

• Close leg of close-then-enter exits is off by slippage·mintick (one extra tick of adverse slip). Wrong exit price → wrong PnL on any strategy using flat-issued opposite brackets with non-zero slippage.
• Entry remainder leg (:546) is correct (single slip).
• Not caught today: every validation probe runs slippage=0.

Probe / test design

1. Engine unit test (deterministic, no TV needed) — fastest confirmation.

• Subclass harness; enter a LONG at known price; arm a flat-issued SHORT bracket entry that fires next bar; set slippage_ = K, mintick = m.
• Assert the emitted close trade's exit_price == apply_slippage(raw, is_buy) once (i.e. raw − K·m for a long-close sell), NOT raw − 2·K·m.
• Add as tests/test_close_opposite_slippage.cpp.

2. Corpus parity probe (TV ground truth).

• Strategy that holds a position and uses a flat-issued opposite bracket entry to close (close_only_opposite path), slippage = N (>0), on ETH-USDT 15m.
• inputs.json: strategy_set_override slippage = N (note: must use a harness that wires strategy_set_override, or bake slippage=N in the strategy() decl).
• Export TV trade list; verify. Pre-fix: close-leg exit prices diverge by exactly N·mintick. Post-fix: parity.
• Also add a slippage>0 variant of an existing close_only_opposite probe (80–87) to guard the regression generally.

Suggested fix

Honor the helper contract — pass the raw fill_price (drop the pre-slip):

// close_opposite_then_enter
execute_partial_exit_qty(fill_price, close_qty);   // was: apply_slippage(...) then pass exit_fill

Then re-run unit test + a slippage>0 corpus probe to confirm single-slip parity. Verify no other caller relies on the pre-slipped value.

Note

Flagged in docs/cheatsheet-runtime-and-execution.md Part 3 ("Possible bug").