Test strategy.risk.max_drawdown at closing fills, as TradingView does - #814

Refs #804

Limits of this change (put under the description of the change)

What it changes: strategy.risk.max_drawdown is now tested at the order fills that close units, on the marked equity at that fill's price (open positions included), against a peak that starts at the equity of the first fill and is raised by the marked equity at fills that close units. A breach closes every open position at that fill's price (exit text "Close Position (Max Drawdown)", empty exit id), cancels the pending orders and blocks later entries. Before, the drawdown was measured once per bar close against a peak made of bar-close equity, and a breach only refused new entries: it closed nothing, and it also refused entries that TradingView lets through. The bar-close measurement and the overflow check of an enormous percentage are kept as they were; only the latch that refused entries at the bar close is removed. The closing leg of a reversing entry is a closing fill, so the new position of the reversal is closed at once when its closing leg breaches the limit. A same-side entry fill is not a test.

What was measured against TradingView (NASDAQ:AAPL 15-minute bars of two sessions, fixed quantities, no commission, no process_orders_on_close, a limit of 0.5 percent of equity, initial capital 10000, pyramiding 3):

1. No test between fills: a long held 1.77 percent under water for hours, closed by the script at a small loss, did not halt.
2. A closing fill at a realised loss above the limit closes every other open position at that fill's price, and later entries never happen.
3. The marked equity counts: a FIFO close whose own loss is below the limit, while the rest of the book is marked far below, closes the whole book at that fill.
4. The closing leg of a reversing entry is tested: the new position is closed at once at the same price.
5. A same-side entry fills while an untested breach exists, and a later same-side entry of the same session fills when it fits the account.
6. A closing fill that realises a profit is tested.
7. The peak is not made of bar-close marks (a drop of 148.75 from the highest bar-close mark at a closing fill did not halt) and not of the equity at an opening fill (a drop of 59.50 from the equity at an opening fill, at a closing fill that left the equity below the start by 2.98, did not halt).
8. A closing fill that realises a profit raises the peak: a long closed at a profit of 94.20 (equity 10094.20 at that fill), then a closing fill 68.00 lower (the equity 10026.20, still 26.20 above the initial capital, the limit 50.47) closed the rest of the book at that fill, and later entries never happened. A reference that stays at the initial capital would not have halted there.
9. The peak counts the open profit of the units still held: a long closed at a profit of 94.20 while a second long held an open profit of 75.36 (marked equity 10169.56); at the next closing fill the marked equity, 10093.06, stood 76.50 below it (the limit 50.85) and the rest of the book was closed at that fill. A reference made of realised equity only (10094.20) would not have halted there.
   Each point is a test scenario on synthetic bars in tests/test_adapter_drawdown_fill_halt.cpp, with the arithmetic in its comment and the digest of its tape; the two pinned scenarios of tests/test_adapter_risk_relower.cpp and the scenario of tests/test_engine_risk_l4a.cpp that this behaviour changes name the tape next to the edit. Each scenario fails on main; the scenarios of points 8 and 9 also fail when the peak is never raised, and the scenario of point 9 also fails when the peak is raised by realised equity only.

What it does not change and does not claim:

1. Not measured: process_orders_on_close, the bar magnifier and calc_on_order_fills; margin-call fills as closing fills; stop and limit fills inside a bar; a short as the first position; the money (non-percent) form of the limit; a commission (the peak starts at the equity of the first fill, which a commission lowers); other symbols and timeframes; other limit values.
2. The peak: whether a closing fill at a loss can raise it is not measured, and neither is the base of the percentage (the peak, the initial capital or the equity at the fill: the drops of the tapes are above all three limits).
3. The cancel of pending orders is not observable on a tape (a later entry is blocked anyway); the test shows that a resting order of the scenario does not fill after the halt, which the entry block alone also explains.
4. The block on later entries lasts to the end of the run in the engine; the tapes show it to the end of the next session only.
5. In one tape a later same-side entry of the same session was absent where the account was short of funds (a twin with a smaller first position let it fill); the reason is not measured and no scenario asserts that entry.
6. No scenario separates the cancel of pending orders from the entry block, pins the order of the latch against the forced close, or sits on the boundary of the comparison (a drawdown exactly equal to the limit).

Known items: this change adds lines to src/source/pine_adapter.cpp, so the documentation-anchor check (scripts/check_doc_anchors.py) reports the anchors below them on this pull request; they are re-anchored once, by one commit on the release branch after the last member pull request. The same commit corrects the statements that describe the previous rule: docs/design/native-feature-parity.md row 541 (a drawdown measured once per script bar at the close mark by update_risk_state against a peak of close marks; a halted strategy that still reverses) and its row 106 and docs/coverage.md line 264, which name update_risk_state as the enforcement point.

🤖 Generated with Claude Code