Bind an exit without from_entry to the market entry placed before it in the same execution - #769

Summary

A strategy.exit given in ticks (profit, loss) and called without from_entry that follows a market entry in the same execution now arms when that entry fills, as the exit with an explicit from_entry already does. Before, its legs took effect with the next call of strategy.exit (one bar late) or, when it was called only on the signal bar, never. TradingView binds such an exit to the pending entry, so the legs can fill on the entry's own bar. Refs #694 and the report in pineforge-codegen-oss#295 (item 5).

What changed

• A relative exit (profit and loss in ticks) that cannot be priced until its entry fills is kept as a pending definition under its exit id and from_entry. materialize_relative_exits, which turns the definitions of an entry into live brackets when the entry fills, only took definitions whose from_entry equals the opening entry. It now also takes the definitions with an empty from_entry when the opening is a market opening (no limit and no stop in the opening snapshot) and the run does not use calc_on_order_fills, after the named ones and after the adoptability check, and re-runs exit() with their profit and loss ticks (the named ones keep passing NaN, as before).
• A definition remembers whether the call was void under the adapter's existing rule (an exit without from_entry called with no position and no working entry order, not under calc_on_order_fills). A void definition is never adopted. The flag lives on the definition because the shadow rows of the calling execution are cleared at the next bar, before the entry fills.
• The state digest folds the flag only when it is set. The flag is set exactly by calls of the void kind, so a run that never makes such a call keeps its digest, and a run that makes one gets a different digest with identical trades.
• CHANGELOG.md has one bullet. The source allocation manifest gets one entry for the new local list (ceiling 1403 to 1404) and its site inventory the three records the gate proposes; the count pin moves by one. 118 documentation citations that the added lines moved were re-anchored.

Tests

• tests/test_exit_without_from_entry_same_execution.cpp (public, synthetic bars, calc_on_order_fills off, five scenarios): an explicit from_entry every bar, an unnamed exit every bar, an unnamed exit on the signal bar only, an unnamed exit called before the entry statement (void, unchanged), and an explicit exit on the signal bar only. The expected trade is computed in the test from the feed and the documented rule, not from the engine. Before the change the second and third scenarios fail (the exit lands on bar 2 at 102.25 instead of bar 1 at 102.00, or never); after it all five pass.
• Nine existing rows around exits and exit tapes (per-entry, same-side, cross-side, bracket re-lowering, exit binding, margin open prints and rules, the fill-bar calc_on_order_fills exits, the exit queue) ran together with the new test, with the private tapes overlaid: before the change the nine passed and the new test failed; after it all ten pass. The private tapes are not part of the public CI.
• The reporter's script on a one-hour BTCUSDT feed: 32 trades and none closed on its entry bar before, 33 trades and 9 closed on their entry bar after (the reporter's TradingView list: 33 and 9); the bar 87 entry closes on bar 87 at 43976.07 after, on bar 88 at 44051.10 before.

What is not covered

• Two market entries that fill at the same open with one unnamed exit were run on synthetic bars: both close on the entry bar at the +200 tick level of their common fill, as two named exits do, whether the exit is called once or on every bar. The unnamed definition is consumed at the first market fill, so an entry that fills on a later bar (pyramiding) is armed as before, by a later call of the exit made while a position is held. No TradingView tape pins these shapes, and the change adds no test for them.
• Priced (limit or stop) entries are not changed: the definition is adopted only for a market opening. A stop entry followed by an unnamed exit called only on the signal bar still never arms, while the explicit form does.
• Under calc_on_order_fills nothing changes: the unnamed definitions are not adopted, as before. No tape says what TradingView does for an unnamed relative exit there, so the change does not reach it.
• No performance claim: the change adds a boolean field to a pending definition and a branch in an existing loop at a market fill; nothing was timed.
• No parity sweep of the corpus has been run on this change. It moves trades on purpose for strategies that use this shape.
• The engine's whole test suite (the hosted CI runs it) was not run by the author.
• Citations in comments and data outside the pages the anchor checker scans (136 rows of ci/cr150-source-reading.tsv, 10 of tests/heap_census_report.md, comments in scripts/ci_verify.py and a few test files) cite lines of src/source/pine_adapter.cpp and tests/CMakeLists.txt that this change moved; they were not moved and no test reads them.

🤖 Generated with Claude Code