INT28: wave-J engine integration (TVDEF-DROPS, W3/W3B/W4/W5/W6/W6B/W8, R1-CONSOLIDATE, RUN-HARNESS) - #294
Merged
Conversation
… fund keeps its closing leg Pine v6 now sizes an omitted order at 100% of equity (lane TV-DEFAULTS). A default percent_of_equity (<= 100) pure STOP entry is sized at its snapped level and priced at the signal close; a sell stop sits below the close, so at 100% its short needs more margin than the equity. ab9714b kept such a reversal as its closing leg (pine_strategy_commands.cpp:408-419). R4 slice C (73817c1) split the default-stop arm off the affordability arm and dropped the whole command, so the long was held to the end of the range. TradingView, on synthetic probes written for this lane (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01; tests/fixtures/tvdef_drops): - tdd-r1-stop-reversal-close-only (v6 defaults): both long seeds close at their sell stops (1548, 1640) and no short opens; - tdd-r1-stop-reversal-half (percent 50, declared): the same stops close the long and open the short. The adapter now sets affordability_close_only for an unfunded reversal in the default-stop arm; the fill-time terms already book a close-only entry as the opposite book's close. A flat or same-side unfunded stop is still dropped. Fail-before: tests/test_tvdef_drops_tapes.cpp built against d3b753c's library: "FAIL ... lane.trades == tape.trades" on tdd-r1-stop-reversal-close-only (each long exits at the 02:15 cleanup instead of the stop), "49 passed, 1 failed"; the 50% control passes. After: "50 passed, 0 failed". Population: scripts jaysharma-order-lifecycle-trace and jaysharmaofficial-alphamojo-supertrend-strategy-ha-with-buffer (20 lanes) and job-2910-cleightyp-claytp-sauce-v1-1 (ETHUSDT.P 1D), whose TradingView tapes never open the short their stop reversals name. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes and no durable state is added (the flag already exists and is hashed). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit db077f39d54fafdc9ad7375b0762eb055607372e) (cherry picked from commit 4ba35993e59fafa1a5938662ff5b00e5d5059283) (cherry picked from commit 04ec289d5519910fadf11cf2221191067bcd2f90) INT28 pick note: TVDEF-DROPS 1 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ersal at the open The adapter decides at a bar open whether a live opposite entry is a reversal TradingView declines (declined_reversal_at_open), and then cancels the position's from_entry stops the bar's adverse extreme reaches (defer_declined_reversal_exits_at_adverse). Both priced a candidate without frozen default units at config_.default_qty_value. That is a fixed default's own size, but an explicit qty=1 entry under Pine v6's default of 100 was priced at 100 units, declined, and its from_entry bracket cancelled -- the bracket TradingView scratches at the reversal's open fill. ab9714b scoped the fill-time decline to a frozen default quantity (pine_fills.cpp:5227-5235, frozen_default_qty), which pine_strategy_commands.cpp:596-601 sets only when qty is omitted; an explicit quantity never entered it. Both loops now skip a candidate whose quantity is explicit; a default candidate is priced as before. TradingView (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01; tests/fixtures/tvdef_drops): tdd-r2-explicit-reversal-bracket (v6 default, every entry qty=1) and tdd-r2-explicit-reversal-bracket-fixed1 (fixed 1 declared) are one byte-identical tape: each one-unit reversal opens and its bracket (stop already through at the open) scratches it there. Fail-before, built against db077f39 (the R1 tree): "FAIL ... lane.trades == tape.trades" on tdd-r2-explicit-reversal-bracket (the short-to-long reversal's long rides to the 02:15 cleanup), "111 passed, 1 failed"; the fixed-1 control passes. After: "112 passed, 0 failed". Population: order-market-reversal-resting-bracket-gap-01 (the ETH@15 hard probe of the same shape: exitP90 0.086 under the v6 default). Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes; no state added. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit c53ad8aecb6462482dd005cc1e782fad193d11a4) (cherry picked from commit 30943479884478945668a28cd0c7daa37b883843) (cherry picked from commit 77b3f544bdfa089b5944819e4c22792dd092d315) INT28 pick note: TVDEF-DROPS 2 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…y="" exits A script that re-issues a from_entry="" exit for the side it holds issues it once more on the bar it places a MARKET entry for the other side -- after the entry, still for the held side. TradingView never applies that exit to the side the reversal opens, on its entry bar or later. The kernel binds such an Independent close to the book it first evaluates against. Under a fixed default the adapter batches the market reversal behind the bar's commands, so the exit is evaluated (and bound) first and the kernel cancels it with the held book (OwnerGone). Under a percent or cash default the reversal reaches the kernel first, the exit binds to the NEW position, and its levels -- marketable against the new side -- close it at its entry open. purge_brackets_after_applied_reversal only ever purged from_entry cohorts (and only for a reversal into a long: its guard reads the signed opened_units). on_applied now cancels, at an applied reversing entry in either direction, every still-working from_entry="" exit placed while the reversed position was open (its side, its cycle). A leg the kernel already cancelled with the held book is not among the host's working requests and is left to its terminal receipt -- a cancel of it would record a NotWorking event -- so the fixed default's path is unchanged, event for event. TradingView (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01; tests/fixtures/tvdef_drops), both directions in each probe: - tdd-r3-reversal-voids-global-exit (v6 default: percent 100, with the 100% shorts' entry-bar margin calls): each reversed side exits later by its own exit; - tdd-r3-reversal-voids-global-exit-half (percent 50): the same; - tdd-r3-reversal-voids-global-exit-fixed1 (fixed 1 declared): the same timing; the engine already booked it. Fail-before, built against c53ad8ae (the R1+R2 tree): "FAIL ... lane.trades == tape.trades" on tdd-r3-reversal-voids-global-exit and on its -half (both reversals scratched at their entry opens by the prior side's exit), "221 passed, 2 failed"; the fixed-1 control passes. After: "217 passed, 0 failed". Population: job-2436-williamcleves-double-tap (NSE:NIFTY 15; 518 trades vs TradingView's 517 under the v6 default: seven reversal-born positions closed at their entry opens by "TP1 Short" / "TP1 Long", one of which re-entered). Behaviour for explicitly sized scripts: an entry reversal whose market request reaches the kernel ahead of the bar's re-issued from_entry="" exit (any non-fixed default, or a fixed default outside the batched same-bar scope: pyramiding > 1, process_orders_on_close, calc_on_order_fills, close_entries_rule=ANY, a priced reversal) no longer lets that exit close the new position. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes; no state added. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit d1a39eb3f75169d2cda52ae764ee829c33af2f0d) (cherry picked from commit c5e724ed1a1ad048694c0fe149d648e6a6d0e9f3) (cherry picked from commit 3d12ef218cdf9bfa6ce45263f47ea82b0a4468e1) INT28 pick note: TVDEF-DROPS 3 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ide its levels are on R3's own tapes leave the reversal-bar exit's levels marketable against the new side, so they cannot tell a void exit from one TradingView keeps but never triggers. This tape settles it: on the bar a MARKET entry reverses the held side, a from_entry="" exit is placed once, after the entry, with levels on the correct side for the NEW position -- and it is void as well. TradingView (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01; tests/fixtures/tvdef_drops/tdd-r3b-reversal-voids-right-side-exit, v6 defaults): the reversal-born long (A: stop 1540 / limit 1570) and short (B: stop 1490 / limit 1440) both ride to the cleanup (the 1570 limit the long reaches at 01:30 and the 1440 limit the short reaches at 01:00 never fill); C's exit, placed with an entry from flat, is that position's and closes it at 1640. The adapter's rule at d1a39eb3 books exactly these rows (no change here). Fail-before, built against c53ad8ae (the R1+R2 tree): "FAIL ... lane.trades == tape.trades" on this tape as on R3's two rule tapes (the engine closes the long at 1570 and the short at 1440), "273 passed, 3 failed". After: "267 passed, 0 failed". Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 377cdce974d2f8ec3c5ece849ae7cdf49009cb0c) (cherry picked from commit ed8a6e806bc76c3e8032ff17f58df6bb8bfbc075) (cherry picked from commit 8633d9440282306f054488487780543ece5dd319) INT28 pick note: TVDEF-DROPS 4 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
TradingView reserves an entry's quantity for its strategy.exit orders in the order the exits were created. Re-issuing an exit modifies it in place and keeps its place in that queue; a later-created exit gets only what the earlier ones leave. Of two exits without qty on one entry, only the one created first ever fills (TAIL-TRIAGE family F05: scalpulat0r's tapes show 946 stop exits and no target exit over twelve lanes). compute_exit_reservation counted every other live exit against the one being re-issued, the exits created after it included. A script that issues a stop and then a target every bar keeps both live and unreserved while flat. On the bar an explicit-quantity entry is placed under process_orders_on_close, the pending entry gives them a basis; the stop, re-issued first, saw the target's stale full reservation, was refused (exit_cancel_bracket), and the target took the whole position. The engine then filled targets TradingView never fills and never fired the stops. compute_exit_reservation now skips exit families queued after the one being re-issued, and reconcile_deferred_exit_reservations hands the position out in queue order. exit_queue_rank reads the queue place from state the adapter already keeps and hashes: the first row of the exit's replacement chain (chain_origin_sequence, which remember() carries across re-issues), a staged leg's predecessor, or the source_sequence a staged leg retained. A new exit still yields to every exit already there, and a strategy.close reservation still counts in full. TradingView (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01; tests/fixtures/exit_queue, fixed 1 per long, five cells): - rule tapes the engine missed: w3f05-s5 (stop A then target B every bar, process_orders_on_close, qty=1: only A fills, though B is reached first in four cells), -s6 (B then A: only B), -s8 (A 50%, B full: B gets the other half), -s8b (A full, B 50%: B nothing), -s13 (both in one oca_name: unchanged); - controls it already booked: -s1/-s2 (s5/s6 without process_orders_on_close), -s7 (50% each), -s10 (B created first on the entry bar keeps the position although A is issued first on every later bar), -s14 (A issued once), -s15 (a cancelled A issued again is new, behind B), -s16 (an exit's place does not carry into a later trade), -s18/-s21 (an exit issued before its entry exists is void), -s20 (one issued while its limit entry rests is the entry's). tests/test_exit_queue_tapes.cpp replays every tape through the Pine adapter. Fail-before, against 8633d944's library: "564 passed, 5 failed" (s5, s6, s8, s8b, s13 differ from the tape); after: "569 passed, 0 failed". Population (Cloud Run, against exp-r5-tvdd-measure-20260927): the twelve scalpulat0r-0dte-scalper-v4 lanes move moderate -> excellent (trade counts unchanged, every trade now the tape's). Their exits name their entries, so of the lane's rules only this one acts there. exp-w3-eng-exit-alloc-c1-20260927 (this rule with a first, broader form of the next) moved these twelve and no other of the 62 F05/F02 probes; the 40 seeded excellent controls of the same lanes stay byte-identical in every experiment of the lane. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes; no state added (the queue place is read from chain_origin_sequence / source_sequence, both already hashed). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 2b259623a488e582fb2853882e0817d0a52d9337) INT28 pick note: W3-ENG-EXIT-ALLOC 1 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An exit with from_entry="" armed before its position opens (issued on the entry bar under process_orders_on_close, or while flat) queues for the position like a named one: TradingView fills only the one created first. The opening fill reconciled the reservations of named cohorts only, so two such exits each kept a dynamic reservation of the whole position and the kernel filled whichever level the market reached first. At an opening, on_applied now orders the global exits still unreserved by their place in the queue (exit_queue_rank) and cancels every one behind an exit that claims the whole position (no qty, qty_percent 100). The one in front keeps its dynamic reservation, as its later re-issues do (the unchanged bracket path); a new exit yields to it as before. TradingView (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01): tests/fixtures/exit_queue/w3f05-s11-pooc-qty-any-entry, s5 with both exits from_entry="": only stop A fills (the rows are s5's, byte for byte). Fail-before, the test at this commit against the previous commit's library: "603 passed, 1 failed" (s11 differs from the tape); after: "604 passed, 0 failed". The rule acts in no run of the three behaviour batteries (test_adapter_quiet_bar, test_adapter_live_state_equivalence, test_publication_witness; an instrumented copy of the adapter counted it). Population: exp-w3-eng-exit-alloc-c4-20260928 (the lane's final tree without this rule) and exp-w3-eng-exit-alloc-c3-20260928 (with it) book byte-identical trades on all 62 F05/F02 probes and the 40 controls. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes; no state added. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 110b270938afb719bd8e4e6bc7e6f0471e8b37c7) INT28 pick note: W3-ENG-EXIT-ALLOC 2 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… too Lane TVDEF-DROPS' R3 cancels, at an applied entry reversal, every from_entry="" exit placed while the reversed position was open and still working on the book: TradingView never applies it to the side the reversal opens. When a second entry of the reversal bar is still pending, the reversal's fill first parks that exit's legs -- marketable against the new side -- in delayed_market_orders_ for the next open, their requests already cancelled. R3 did not see them there: the adapter released them at the next open and closed the new position with the reversed side's levels. Meanwhile they also stood as reservations against the new side's own exit. The R3 block now also drops such parked legs (and staged bracket legs of the same kind) from their queues. They left the book when they were parked, so dropping them records no kernel event. TradingView (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01): tests/fixtures/exit_queue/w3f05-r1-two-parent-reversal-global-exit (40% of equity; a long re-issues exit XL, from_entry="", at close -/+ 20 every bar; at 00:30 two short entries S1 and S2 and XL again; the short re-issues XS at avg -/+ 15): every short exits by XS or the cleanup, never by XL. Population witness: job-2436-williamcleves-double-tap on OANDA:XAUUSD 15, 2026-01-16 15:30 UTC (PUT_SCOUT and PUT_BOS both placed at 15:15; TP1 Long's parked legs closed the short at 16:00 open, where TradingView's TP1 Short stop fills at 4579.32). Fail-before, the test at this commit against the previous commit's library: "684 passed, 1 failed" (r1 differs from the tape); after: "673 passed, 0 failed". lab tv tapes of tests/fixtures/tvdef_drops still replay (267 passed). Population (Cloud Run, against exp-r5-tvdd-measure-20260927): exp-w3-eng-exit-alloc-c4-20260928 (the first rule and this one) and exp-w3-eng-exit-alloc-c3-20260928 (the lane's final tree) move four job-2436-williamcleves-double-tap lanes that exp-w3-eng-exit-alloc-c1-20260927 (the first rule and a first form of the second, without this one) left alone: OANDA:XAUUSD 15 strong -> excellent (the witness above: that short now exits at 4579.316; 4 unmatched pairs -> 0), NYSE:F 1D strong -> excellent, NASDAQ:AAPL 1D strong at match 96.3 -> 98.1 %, OANDA:EURUSD 15 excellent with TradingView's 651 trades (650 before). No metric of the 62 F05/F02 probes worsens; the 40 controls stay byte-identical. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes; no state added. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 977da92e464bbf74e3f875ad2c5694ddea16ab91) INT28 pick note: W3-ENG-EXIT-ALLOC 3 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ills oldest first
Under the default FIFO close_entries_rule TradingView sizes a
strategy.close(id) without qty from a ledger of the units entered under the
id that no fill has booked yet. The grid bots of population family F02
(TAIL-TRIAGE, 43 probes: pyramiding up to 200, one strategy.close call site
in a loop over the grid levels) read that ledger on every take-profit, and
TradingView keeps it by three rules, pinned by the tapes in
tests/fixtures/grid_close:
- one call site places one order per bar; a later call of the site re-sizes
that order for its own id and comment, but the fill books against the id of
the site's first call on the bar (w3f02-x2, w3bf02-a2);
- a fill books against its id's unbooked units, oldest first, and what they
cannot cover spills over every id's unbooked units in the order the entries
filled -- not by id name, and a re-entered id's new units behind every older
entry's (w3bf02-a1, -b1);
- a call whose id has no unbooked units places nothing and leaves the site's
order as it was (w3bf02-c1, -c2, -c3).
The adapter kept a per-id ledger but erased the first id's units whole and
held the survivor's fill as a standing reservation against the other ids
instead of booking it, with a carried "first target" for exact two-call
batches. The units the first id could not cover stayed on the books:
close("L38") closed 0.19 where TradingView closes 0.18 (w3f02-x2) or nothing
(w3f02-x1), and a loop closing D then C booked C (w3bf02-a2). Replaying the
43 probes' own call logs through the three rules reproduces every close bar
of TradingView's tapes before the first script-state divergence; the
reservation model missed 120 of them.
The ledger now keeps, beside the per-id totals, the same units by opening
fill in fill order (close_ledger_records_; adjacent records of one id
merge). enqueue_pooc_fifo_close sizes a call from the id's ledger against the
position less the bar's other close orders, as before, and skips a call with
nothing to close; the site keeps its first id. The fill books through
book_close_ledger. The reservation and first-target maps, the site's carry
and deferred-cleanup fields and close_reserved_other_units are gone; their
fixture reads answer none. The state fold keeps its layout: the records fold
only when there are two or more (one record is what the per-id map already
folds), the four retired maps fold as the empty maps they were outside a
multi-call site, and the five retired PlacementSnapshot close fields stay at
their defaults.
tests/test_grid_close_tapes.cpp replays 18 tapes (lane W3's w3f02-x1, -x2
and -g2 .. -g9, this lane's w3bf02-a1 .. -e2) and requires every trade,
exit signal included. Against 75fc4b90's library: "1046 passed, 3 failed"
(x2, x1, a2). After: "1046 passed, 0 failed".
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 5334a42db7c208b66aa21c79272dfbbb31944d3b)
INT28 pick note: W3B-ENG-GRID 1
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
TradingView's close ledger (previous commit) sums to the position: every
fill that reduces it books units, not only a strategy.close(id) without qty.
A fill books against the entry it names -- strategy.close(id, qty=q) against
id (tape w3bf02-f3: close("B", qty=0.1) takes A's lot physically, yet
close("B") then closes B's 0.1 left and close("A") A's whole 0.1) and a
strategy.exit against its from_entry -- and spills what that entry cannot
cover over the oldest entries' units. A fill that names no entry books the
oldest entries' units: a strategy.order sell (w3bf02-f1: after it
close("A") is void) and a margin call. The margin call's booking has no
synthetic tape; replaying the F02 probes' call logs, only oldest-first
booking reproduces TradingView's closes after the one margin call inside an
evaluable window (3commas-eth-grid ETHUSDT.P 1D, 2026-03-09): leaving the
ledger unbooked misses three later close bars, booking the liquidated lots'
own ids six.
The adapter booked only process_orders_on_close strategy.close(id) sites.
observe_close_ledger now books every fill that leaves its side open (a
reversal or a flat clears the ledger first): a Close row against the site's
first id or its own id, an exit leg against its from_entry, anything else --
an order, a margin call, a risk action, a fill with no placement row --
against nothing. Under the ANY rule the ledger sizes nothing and books
nothing (a close is bound to its id's own entries there), and it keeps no
records.
tests/test_grid_close_tapes.cpp gains the two rows. On the previous
commit's library: "1104 passed, 2 failed" (f1, f3). After: "1100 passed,
0 failed". The host's strategy_order passes no comment (nor does the
generated call), so f1's TRIM exit compares unsigned.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 93c16ce8c0a684a50c9f074f6c0e1e9855a582d2)
INT28 pick note: W3B-ENG-GRID 2
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…se ledger
W3-ENG-EXIT-ALLOC's open finding g1. Without process_orders_on_close
TradingView sizes a strategy.close(id) without qty from the same close
ledger as with it: the units entered under the id that no fill has booked,
filled at the next open, whether or not the FIFO rule has already closed the
lots that carried the id. Tape w3f02-g1: after close("B") takes A's lot,
close("A") still closes 1. The first-call booking holds there too (w3bf02-d1:
the loop's order fills 0.22 at the next open, booked against L45, and
close("L38") closes 0.18), and so do the spill order (w3bf02-d2) and an
exit's booking against its from_entry (w3bf02-f2: the exit for B closes A's
lot but books B's units, so close("A") closes A's 0.1 and close("B")
nothing).
The adapter sized the call from the lots still carrying the id
(cohort_exposure_for) and dropped it when the FIFO rule had closed them all
(pine_adapter.cpp, "compute_close_target_qty drops close when target unclosed
quantity is zero"). A default FIFO close (no qty, no qty_percent, not
immediately, not the ANY rule) now sizes from the ledger, capped at the
position, and is dropped only when the ledger holds nothing for its id. The
ledger is a running sum of fills: where it differs from the id's own live
lots by no more than kQtyEpsilon, the lots are the quantity, so a close of
every lot of the id leaves no rounding dust (a streamed run of
test_adapter_host_view_memo closed 3 ULP less otherwise). A later call of the
same call site on the bar replaces the site's order, as before, and now
carries the site's first id (close_first_id, set only when it differs from
the call's own id), which the fill books against. Under
process_orders_on_close the batched path already sized from the ledger; its
other routes keep their basis.
tests/test_grid_close_tapes.cpp gains the four rows. On the previous
commit's library: "1306 passed, 4 failed" (d1, d2, f2, g1). After: "1310
passed, 0 failed".
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 34656aa3071db7eec841893b3f692f1c524ba03e)
INT28 pick note: W3B-ENG-GRID 3
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
W3-ENG-EXIT-ALLOC's open finding (tapes s17, s19; s12 interacts). TradingView voids a strategy.exit(from_entry = X) when the call runs while X has neither an open trade nor an entry order waiting to fill: the exit does not wait for a later entry of that id. w3f05-s17 (process_orders_on_close) and -s19 (next open) issue target B once, on the bar before each entry, and TradingView never fills it; lane W3's s18 and s21 already showed the same on the entry's own bar, before the entry call, and its s20 that an exit issued while its limit entry rests is the entry's when it fills. A new synthetic tape, w3bf05-v1, shows the void call issued between two trades of L never filling in the next one, where the engine filled it (04-10). The adapter kept such an exit in the book, cohort-bound, and filled it against the next position X opened (B at 00:15 on three of s17's five days). The same flat-time exits cost s12 its stops: B, re-issued while flat after the previous trade, kept a queue place ahead of the next trade's stop A, issued first on the entry bar (lane W3's rule A queues exits by creation), so A never fired inside the hour on 04-09 and B took 04-14 at 00:15. exit() now marks a call void (PlacementSnapshot::void_issue) when a named from_entry has no live lot -- read off the book too (open_lot_of), since a reversal by an entry of the reversed side's own id clears that id's cohort facts after recording its new lot -- no unbooked close-ledger units (under FIFO), no entry order pending anywhere the adapter holds one (entry_order_pending) and no exit of the pair issued while the entry rested or was open (standing_exit). The void call's request is still placed, so every later request keeps its incarnation (a script that issues both sides' exits on every bar exports the same entry incarnations), but it acts on nothing: it reserves nothing (compute_exit_reservation, the reconciliation and the named-exit projection skip it), holds no place in the exit queue (exit_queue_rank, remember and the staged paths start a re-issue made once the entry exists afresh, without the void request's leg lifecycle), rests for no margin checkpoint, is passed over by the post-calculation bracket execution, the delayed open release and the declined-reversal stop scan, stays bound to its entry's cohort, and the entry's opening withdraws it wherever it waits (withdraw_void_exits). Where the process_orders_on_close close pass or the recalculation's same-close fill takes it up, the close it places is withdrawn instead of executed (execute_or_withdraw_close): that pass used to flatten the book on the other side's void exit (a short closed by an exit of an entry that never existed, in the publication witness's storms), and it still places the close, so the numbering stays. A call made once the entry exists that leaves a void request unchanged takes it over, queued from that call (unvoid_exit). A void relative exit leaves nothing for the fill to resolve. The state hash folds void_issue only when set. An all-NaN exit keeps its cancel of a prior bracket; exits for every entry (from_entry="") are not touched. v1 also shows TradingView keeping an exit issued while its limit entry rested when strategy.cancel cancels that entry, for the next entry of that id, and ignoring a void call over it (04-08, 04-09). The adapter withdraws such an exit with its entry: an open finding outside this rule, recorded in tests/fixtures/exit_queue/README.md; until it is fixed a void call over an exit the adapter still holds for its pair re-issues it as before. v1's row compares the engine's trades from 04-10 on. tests/test_exit_queue_tapes.cpp gains the four rows. On the previous commit's library: "841 passed, 4 failed" (s12, s17, s19, v1). After: "845 passed, 0 failed". tests/test_flat_coof_exit_tapes.cpp's five tapes of an exit placed for an id with no order yet (lane PAR-ORDERS-2), whose recalculation rows it recorded as divergences, now book TradingView's rows exactly; the recorded marks go. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 9568c524b7afb56331d49fc975eaac46d6960c89) INT28 pick note: W3B-ENG-GRID 4 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…e leaves Lane W3's review finding: rule B (110b2709) queues the exits for every entry (from_entry="") at the opening only by cancelling the unreserved ones behind a full one. A full exit behind partial ones kept its dynamic reservation of the whole position. TradingView queues them like a named entry's exits (w3bf05-g8, lane W3's s8 with both exits from_entry=""): behind the 50% stop A, target B takes the other half and never the whole position -- its rows are s8's, byte for byte. The engine closed all of it by B on four of five days. When an unreserved full global exit has partial ones ahead of it in the queue, the opening now hands the position out through reconcile_deferred_exit_reservations with an empty from_entry, the path a named entry's exits take (rule A); every other shape keeps rule B's cancel as it was (a reserved exit neither triggers nor suffers it). Two more review findings show no trade effect on TradingView's tapes and stay open, pinned by rows: w3bf05-p1 (the reservation caps ignore queue order: a front exit raising its qty_percent still takes the whole position the bar it is raised) and w3bf05-r1c (rule C leaves the recalculation queue alone: lane W3's r1 under calc_on_order_fills books TradingView's rows). The fourth -- rule B cancelling a request the kernel already ended -- changes a hash, not a trade; tests/fixtures/exit_queue/README.md records it. tests/test_exit_queue_tapes.cpp gains the three rows. On the previous commit's library: "1008 passed, 1 failed" (g8). After: "1021 passed, 0 failed". Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 3bce6bf342f88e2a62064793f174866bb3ac9120) INT28 pick note: W3B-ENG-GRID 5 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…gView pins
The close model of this lane (its F02 and g1 commits) retires the
adapter's close reservation maps: a call site's order books against its
first call's id and spills over the oldest ledger records, and every
reducing fill books the ledger. test_integration_l4d (the frozen twin of
test_integration) pinned that reservation model's own bookkeeping in its
L4d close-replacement probes, several of whose values TradingView
contradicts. Six synthetic lab tv --no-note tapes of those probes' shapes
(tests/fixtures/grid_close, w3bl-*) pin TradingView's rows:
- w3bl-s6: one site closes F7 then L7, F15 then L15, then L3 then L4: the
last order closes L4's whole 0.4159 (the reservation model capped it at
0.3560; the twin's debt 0.5629 -> 0.6228, admitted 0.3560 -> 0.4159, final
position 0.8982 -> 0.8383);
- w3bl-s7: a sell of 1 uses A up, so two sites' closes of A are void and the
second site's close of B takes the book (the model rejected it: no exit,
no debt; site B / REJECTED_FULL_B, two trades);
- w3bl-s8: a later close of D closes D's 3 (the model capped it at 2);
- w3bl-s9: after a sell of 1, a close of D closes D's 4 and flattens the
book (the model closed 3);
- w3bl-s10a, -s10b: two sites on B's re-entries in either order, a sell of
2, one site closing B, C then D: B's and C's calls are void and D's order
closes D's 4, in two trades of 3 and 1 (the model closed 3, or 3.4).
tests/test_grid_close_tapes.cpp gains the six rows and the fixture README
their table. The twin's probes keep only the ledger view (its
reservation-map fixture accessors, and l4d_native_route_guard's, go with
the maps); its thirteen affected tests are renamed for what they now pin and
their assertions follow the tapes. tests/twin_parity_inventory.json extends
test_integration's observable rewrite (569 twin assertions, digest
re-derived, reason citing the tapes) and scripts/check_twin_parity.py
re-pins INVENTORY_SHA256; tests/test_integration.cpp, which the twin had
already left behind, stays as it is.
On the base 75fc4b90's library the six rows fail, beside the nine rows of
the lane's F02 and g1 commits ("1562 passed, 15 failed"); on this tree:
"1556 passed, 0 failed". test_integration_l4d: "1178 passed, 0 failed".
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 92348938d568d376b8019ea800ee6aaf14aee8a8)
INT28 pick note: W3B-ENG-GRID 6
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The lane's void rule (an exit whose entry neither exists nor rests is void, TradingView's tapes w3f05-s17 / -s19 on top of lane W3's s18 / s21) reaches two frozen twins whose scenarios issue an exit before its entry's call, while the entry neither exists nor rests: - test_dual_entry_placement_sizing_l4b, test_MM_pair_defers_percent_exit_ reconciliation_until_net: the pair's 40% and 60% exits for LAY-L are issued between LAY-S's and LAY-L's calls, so nothing rests for them after the pair (1.0 and 1.5 -> 0 and 0); its legacy copy tests/test_dual_entry_placement_sizing.cpp follows; - test_prearmed_exit_path_cursor_l4c, the FreshChild book of check_missing_provenance_is_inert: X is issued after the parent's cancel and before the fresh parent's call, so the fresh parent's trade has no X to close it (one closed trade -> none); the MissingParentCancel and PostCancelDoubleReissue books keep theirs. tests/twin_parity_inventory.json records both as observable rewrites (their twin assertion counts and digests, the reason citing the tapes) and scripts/check_twin_parity.py re-pins INVENTORY_SHA256: "155 tests, 8298 base assertions, ..., 404 pinned observable rewrites, ... OK". On this tree: "178 passed, 0 failed" and "553 passed, 0 failed"; on the previous commit's tree each fails its corrected assertions (2 and 2). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit effb648d0e396673ffd91f5ad4ae5b39ca23aae7) INT28 pick note: W3B-ENG-GRID 7 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…r a declined reversal and a margin-call revival At Pine v6's defaults a MARKET reversal the next open cannot fund is declined there, and its same-bar reversal pair holds the reversed side's brackets dormant until a margin-call slice revives them. The adapter did this only for a bracket naming the held entry: follows_same_bar_declined_reversal returned early for a from_entry="" exit, suspend_brackets_for_reversal skipped it, and revive_brackets_after_margin required a named cohort's exposure. A global exit names every entry, so it is always the held position's; TradingView books a global exit exactly as it books the named one. TradingView, on synthetic probes written for this lane (lab tv --no-note; tests/fixtures/margin_v6): - w5-mr-global and w5-mr-once-global (BINANCE:ETHUSDT.P 15, v6 defaults): shorts reversed on their fill bar by a default long, with a from_entry="" exit already through at the next open -- the same file as the named-exit controls w5-mr-named and w5-mr-once-named; - w5-rv-global-held (NYSE:F 15, the V19D-P1 shape): the stop exit placed from_entry="" while short is revived by the 05-02 09:30 margin call and closes the rest (12 @10.39 called, 928 @10.39 through X) -- the same file as the named control w5-rv-named-control. Fail-before (tests/test_margin_v6_tapes.cpp against the base 8633d944): w5-mr-global, w5-mr-once-global and w5-rv-global-held fail; the three named controls pass. After: every row passes. The test row is registered here, with the release floor 718 and the PR registration floors 699/699/708 counted with ctest -N on this lane's tree (wave J's two tape rows: TVDEF-DROPS' and this one). Population (Cloud Run, exp-w5-eng-margin-v6-s4-20260928 against exp-r5-tvdd-measure-20260927): job-2436-williamcleves-double-tap (ETHUSDT.P 15) strong -> excellent. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes and no durable state is added. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 5aac65a65ffe3cbaba10ba31976797c2fe3bfec3) INT28: the rule, the tape row and its fixtures apply unchanged (each file's patch is byte-identical to the lane's). Its floor hunks (docs/ci.md, scripts/ci_verify.py, scripts/test_ci_verify.py: release 718, PR registration 699/699/708 on the lane's tree) are not applied; INT28 recounts the floors once on the integrated tree. tests/CMakeLists.txt: git's clean merge put this lane's fixture if() inside W3-ENG-EXIT-ALLOC's block (its endif() fell below the new block); each block is closed on its own. INT28 (with W3B-ENG-GRID picked before it): composed with W3B's tape w3bf05-r1c (W3 rule C under calc_on_order_fills, tests/fixtures/exit_queue). There S1 reverses the long and S2, a second short MARKET entry of the bar, is rejected at the same open; is_declined_market_reversal() reads S2's placement (a short over a long) as a declined reversal, and this rule's from_entry="" extension held the short's own exit XS, placed in the recalculation after S1 filled, so XS missed its entry-bar limit (1020 passed, 1 failed). A from_entry="" exit placed under the side the rejected entry would open is that side's, not the held position's: follows_same_bar_declined_reversal() and suspend_brackets_for_reversal() leave it be, as bracket_belongs_to_reversal() already leaves a named exit of that side. The w5-mr-* and w5-rv-* exits, placed under the held side, are held as before: test_margin_v6_tapes 3081/0, test_exit_queue_tapes 1021/0. INT28 pick note: W5-ENG-MARGIN-V6 1 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…its combined margin A default percent_of_equity (<= 100) MARKET add to a held position of its own side reached the kernel unjudged: the placement affordability rule covered explicit, fixed, cash and above-100 % sizes only. Where an exit or a close placed on the same bar took the next open first, the add then opened a new position from flat. TradingView drops, when it is placed, an add whose margin on top of the held units exceeds the equity. TradingView, on synthetic probes (lab tv --no-note, BINANCE:ETHUSDT.P 15; tests/fixtures/margin_v6): - w5-m1-addexit-p100 (100 %, pyramiding 10): every add, placed with an exit the next open is already through, is dropped; the exit alone fills; - w5-m1-addexit-p60 (60 %, the control): 60 % + 60 % exceeds the equity too; - w5-m1-addclose (100 %, pyramiding 2): an add followed by strategy.close of the held id, or by strategy.close_all, never opens. entry() now drops such an add inside the pyramiding cap (add_margin_exceeds_equity: the held units plus its own at the signal close's tick against the placement equity, with the one-lot slack the fill-time add arm keeps -- test_reversal_admission_float_guard pin E, a sub-lot shortfall add admitted, still passes). An over-cap all-in add the cap check lets through stays a request, as before; apply_open_market_admission, which keeps such a row only where an earlier opposite command of its batch reverses the book first, now also retires it when that placement judgement failed and the batch holds this bar's market entries and closes alone (the priced, raw and carried controls of test_live_position_market_gross_admission keep their ordinary fills). Judged there, the row keeps its request incarnation, so a script whose over-cap rows never fill books byte-identical output. test_declined_reversal_close_leg_l4c R6 (an all-in add followed by a close of the held id, pyramiding 2) expected the add to re-open from flat after the close, LONG 100: expectation corrected to TradingView's w5-m1-addclose -- the close fires and the book is flat. Fail-before (the MR tree): w5-m1-addexit-p100 and w5-m1-addclose fail, the p60 control passes. After: all three pass. Population (Cloud Run, against exp-r5-tvdd-measure-20260927): masayanfx-scalping-london-new-york-strategy on all ten lanes moderate/strong -> excellent (pyramiding 10: the add behind the same bar's exit), van007trader-quick-reversal-candles-dyna on four lanes strong -> excellent (a default over-cap add paired with the reversal before it: TradingView never reverses back). Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes and no durable state is added. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 79782e9268bd74e09e374830d73ac10ef519be18) INT28 pick note: W5-ENG-MARGIN-V6 2 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…d at its high before the script ab9714b checked a commissioned or slipped short carried under process_orders_on_close only after the script had run at the close (on_bar_close), skipped the check once a slice had been taken at the bar's open, and deferred it behind a close-time market order to the post-fill book. TradingView calls the carried short at the bar's high before the script runs: the script reads the called book and equity, and its close fills meet what the call left. TradingView, on synthetic probes (lab tv --no-note, BINANCE:ETHUSDT.P 15, process_orders_on_close, commission 0.05 %, slippage 1; tests/fixtures/margin_v6): - w5-m2-pooc-short-comm: a call at the next open and another at the high of the same bar; - w5-c1-seen-size: the 12:00 short called at the 12:30 high (1.5312 @1802.01); a reversal on that bar by math.abs(strategy.position_size) opens 53.645, the size after the call; - w5-m2-pooc-carried-close / -none / -reverse: the same call whatever the close does (close_all, nothing, a default reversal sized from the equity after it). The new hook PineExecutionAdapter::on_bar_close_before_script, which PineStrategyHost::on_native_bar calls ahead of the scheduler's script callback, takes that check for the commissioned or slipped carried short (submit_margin_call_slice at the bar's high). A fee-free carried short keeps the kernel's path check and the checkpoint after the script, with its deferral, unchanged. test_adapter_margin_schedule_differential expected, for a carried POOC short with a fee closed by the script on the bar whose high calls it (M7-C, and the closing run of E20 f1), no margin call: the check ran after the close had emptied the book. Expectations corrected to TradingView's w5-m2-pooc-carried-close: 1.7504761904761923 @105 at the calculation point, the kernel's own slice, then the close of the 7.7495238095238079 left. Fail-before (the M1 tree): w5-m2-pooc-short-comm fails. After: it passes. The carried tapes also need the next open's call of M2 (and w5-m2-pooc-carried-reverse the long's call of C2); the test replays them from those commits on. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2); the host only calls the adapter's hook. No kernel unit changes and no durable state is added. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 8577227977b61959d32a4c64f397cf450fcf3224) INT28 pick note: W5-ENG-MARGIN-V6 3 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ts signal close A 100 %-of-equity short that process_orders_on_close fills one tick under its signal close is short of margin by its slipped fill and its entry fee. TradingView sizes that call at the signal close's mark and executes it at the next open (re-sized at the open only when the close's units no longer cover the open's own shortfall), then checks the survivor again at the call's slipped fill. The adapter sized the opening slice at the open alone and took no re-check at the fill. TradingView, on synthetic probes (lab tv --no-note, BINANCE:ETHUSDT.P 15; tests/fixtures/margin_v6): - w5-m2-pooc-short-open (commission 0.05 %, slippage 1): fifteen shorts whose next open is one tick above, at or below the signal close -- each first call sized at the close; - w5-m2-pooc-short-plain: the fee- and slippage-free control; - every first opening call of the 258 such shorts in lab tv tape officialjackofalltrades-caldera-meridian-strategy-joat follows the close's mark; marking at the open alone books 118. on_bar_open keeps the script bar the opening follows (prior_policy_bar) and sizes a short filled at that bar's close from its close (close_margined_units); on_applied re-checks the survivor of such an opening call at its slipped fill (recheck_at_fill) before the remaining path. Fail-before (the C1 tree): w5-m2-pooc-short-open, w5-c1-seen-size, w5-m2-pooc-carried-close and w5-m2-pooc-carried-none fail; the plain control passes. After: every row passes. Population (exp-w5-eng-margin-v6-s4-20260928 against exp-r5-tvdd-measure-20260927): officialjackofalltrades-caldera-meridian- strategy-joat (ETHUSDT.P 15) strong 99.2 -> 99.8, with C1, QP and C2. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes and no durable state is added. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit c4f79213d684550d7468922eb007505a9409e6b3) INT28 pick note: W5-ENG-MARGIN-V6 4 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… took from its entry TradingView sizes a percent strategy.exit on the quantity its entry filled: the slices margin calls took from the position are not taken off that basis. The adapter reserved a percent exit on the position left after the calls. TradingView, on synthetic probes (lab tv --no-note, BINANCE:ETHUSDT.P 15, process_orders_on_close, commission 0.05 %, slippage 1; tests/fixtures/margin_v6): - w5-qp-pooc-reissue: a 62.3911 short called 0.0028 and 1.6712 before its TP1/TP2 (35 % each, re-issued every bar) are placed -- each closes 21.8368, 35 % of the 62.3911; - w5-qp-pooc-once: the control, placed once at the entry (the same file). exit() adds the units closed by margin calls since the position opened (Trade rows labelled kMarginCallLabel on its side, its FIFO entry for a named exit) to the book basis of a qty_percent reservation below 100 %. Fail-before (the M2 tree): w5-qp-pooc-reissue fails, the once control passes. After: both pass. Population (exp-w5-eng-margin-v6-s4-20260928 against exp-r5-tvdd-measure-20260927): officialjackofalltrades-caldera-meridian- strategy-joat's TP rows (ETHUSDT.P 15), with M2. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes and no durable state is added (the basis is read off the host's closed trades). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 39961733ed662a19a2cca8f1b75f2bcd4df416f8) INT28 pick note: W5-ENG-MARGIN-V6 5 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…s called at the next open The close fill that opens, or reverses into, a commissioned process_orders_on_close position was margin-called at the fill: a reversal's closing fee leaves the new 100 % side short of margin. TradingView takes no call there; it calls that book at the next open -- a long for the units the fill's checkpoint would have taken, a short at the close's mark (M2). TradingView, on synthetic probes (lab tv --no-note, BINANCE:ETHUSDT.P 15, process_orders_on_close; tests/fixtures/margin_v6): - w5-c2-pooc-rev-long (commission 0.05 %, slippage 1) and w5-c2-pooc-rev-long-noslip (commission alone): fifteen shorts reversed to longs, each long called at the next open, never at the fill; - w5-c2-pooc-rev-short: fifteen longs reversed to shorts, each called at the next open at the close's mark; - w5-c2-pooc-flat-explicit-long / -short (commission 0.1 %): fifteen explicit-quantity positions of strategy.equity / close each, short of margin by their entry fee, every one called at the next open; - w5-c2-pooc-long-open: the control, a 100 % long from flat, never called; - w5-m2-pooc-carried-reverse: C1's carried short reversed, the long called at the next open. on_applied skips the fill-time opening checkpoint for a commissioned process_orders_on_close close fill that opens the position (called_at_next_open); on_bar_open calls such a full-margin long at the open (submit_margin_call_slice at the open's tick). A short is called there by M2's opening slice. test_margin_call_1x_long_entry_fill_l4a F (a POOC long of explicit 99.95 @100 under a 0.1 % fee) expected the 0.1996 trim on the fill bar, and test_pooc_open_money_event_l4b's fee compatibility scope expected its opening event at the entry close: both now book it at the next open, as w5-c2-pooc-flat-explicit-long does. Fail-before (the QP tree): w5-c2-pooc-rev-long, -rev-long-noslip, -rev-short, -flat-explicit-long, -flat-explicit-short and w5-m2-pooc-carried-reverse fail; the long-open control passes. After: every row passes. Population (exp-w5-eng-margin-v6-s4-20260928 against exp-r5-tvdd-measure-20260927): job-2550-spiritedcheeta2101-quanloki-qqe- system-opposite-signal-only-points (process_orders_on_close, commission 0.05 %, slippage 1) on OANDA:XAUUSD 15 and 1D strong -> excellent, on OANDA:EURUSD 15 94 fewer trades (3553 against TradingView's 3551), with C1. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes and no durable state is added. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 64decc5c57e0e9ae059061e4c3f10cdb7e2d2c08) INT28 pick note: W5-ENG-MARGIN-V6 6 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…t its print A 100 % market short filled one tick under the open (slippage 1) was checked at its slipped fill price, where its fee-short margin is still covered. TradingView sizes the opening call at the print the fill slipped from and executes it there with the call's own slippage; the high then calls the survivor. TradingView, on synthetic probes (lab tv --no-note, BINANCE:ETHUSDT.P 15, commission 0.05 %, slippage 1; tests/fixtures/margin_v6): - w5-qp-mkt-once and w5-qp-mkt-reissue: a short filled at 1611.06 under the 1611.07 open is called 0.0028 @1611.08 there, then 0.4716 at the high; - w5-mk-short-rev: fifteen longs reversed to market shorts at the next open, each short called at its open print first. on_applied's fill-time opening checkpoint takes a slipped market short's call at the print (source_bar_fill_tick of the fill plus the slippage), not at the fill; every other opening keeps the fill's checkpoint. Fail-before (the C2 tree): all three rows fail. After: they pass. Population (exp-w5-eng-margin-v6-s4-20260928 against exp-r5-tvdd-measure-20260927): job-2634-tetradprotocol-tetrad-elektro- community-v1-0 (commission 0.05 %, slippage 3) on ETHUSDT.P 15 weak -> excellent and on OANDA:EURUSD 15 strong -> excellent, job-2764-zhangwenjian810-mgc-wyckoff-second-breakout-v3-candle-rr (commission 0.02 %, slippage 2) on BINANCE:BTCUSDT 15 strong -> excellent. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes and no durable state is added. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 266a28bec5ae383f2805f7219d0d3de85caffc24) INT28 pick note: W5-ENG-MARGIN-V6 7 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ket, buy stop, sell market, sell stop, buy limit, sell limit Two strategy.entry calls placed on one bar from flat, both already marketable, fill at one price, and TradingView fills them in a fixed order of side and kind -- buy market, buy stop, sell market, sell stop, buy limit, sell limit -- one rank in placement order. The later of two opposite calls is one transaction of its own quantity plus the earlier pending MARKET call's; the earlier call trades only its own. So a marketable buy stop placed after a sell market fills first as a buy of both quantities, and the sell then closes part of that lot: TradingView's two rows carrying the buy's signal, which the grader consolidates into one qty-2 trade (TAIL-TRIAGE F10-KI65-DUAL-ENTRY). The engine queued the calls in the order they were placed, apart from the market pair the same-bar batch already materializes buy first and the pure-stop pair defer_open_marketable_sells re-sequences at the open, and it sized a buy stop placed after a sell market as a reversal against it. - PineExecutionAdapter::order_same_point_entries (src/source/pine_adapter.cpp:19780), called after every source evaluation (src/source/pine_strategy_host.cpp:1521): within the flat explicit-quantity family same_bar_market_tx_scope() settles as one broker batch, the bar's entries that are already marketable at the evaluation's price are queued in TradingView's order. Every request from the first one out of order on is re-priced in place (ReplaceOptions::keep_handle, the kernel's V19-D capability), which keeps its handle and ranks it newest; the kernel breaks a tie at one cursor oldest request first. A pure-stop book is left to defer_open_marketable_sells. - entry() (src/source/pine_adapter.cpp:6625): a marketable buy stop that follows a pending sell market on its flat bar is one Transact of its own quantity plus the market's (pending_opposite_market_units, src/source/pine_adapter.cpp:19737), not a reversal sized at its fill. TradingView (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01; tests/fixtures/same_point_entries): - w6-f10a-open-pair: 14 cells (MM/MS/SM/SS/ML/LM/LL/LS/SL x LF/SF, MS-SF-Q), twice each; the engine booked MS-SF, SM-SF, LM-SF, LL-SF, LS-SF and MS-SF-Q sell first. - w6-f10b-limit-class: a limit leg declared first or second, and the Q cells; the engine booked LM-LF, LS-LF, LM-LF-Q and LM-SF-Q in placement order. - w6-f10c-same-side-class (pyramiding 2): a market lot fills before a stop or limit lot of the same side placed before it. Read off the tape only: the scope excludes pyramiding 2, where the engine keeps placement order (README). tests/test_same_point_entries_tapes.cpp replays f10a and f10b under the configuration their generated constructors declare, and reads the order and the buy's quantity off TradingView's rows of every pair tape. Fail-before, built against 8633d944: "FAIL tests/test_same_point_entries_tapes.cpp:376 got == want" on both rows, "765 passed, 2 failed". After: 767 passed. test_dual_entry_placement_sizing R2 (MS-SF-A) had pinned the engine's sell-first decomposition beside TV's net position, noting TradingView's split differs; it now pins TradingView's: the closed trade is a slice of E2's long, closed by E1. A frozen twin assertion, so tests/twin_parity_inventory.json records the observable rewrite with its reason (and the checker its sha). Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2); the kernel is untouched and keep_handle is its generic re-price, no new capability. No state added: the order lives in the kernel's queue and the live-handle roster, both already folded. Population: Cloud Run experiment exp-w6-eng-fill-order-a-20260927 on this lane's final engine tree, 2,166 probes (the 24 of families F10 and F12, 40 seeded excellent controls and the static census of every rule's shape), against exp-r5-tvdd-measure-20260927: pf-probe-ki65-dual-entry-precedence on es1-15, eurusd-15 and xauusd-15 moderate -> excellent (550, 554 and 550 unmatched engine trades -> 0); its nq1-15 row needs the next commit too. The 40 controls are byte-identical. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 443e760598bd2ba2aef7fbc5212e9f55c0320722) INT28: every code, header, test and fixture hunk applies unchanged (each file's patch is byte-identical to the lane's). scripts/check_twin_parity.py: INVENTORY_SHA256 is the merged inventory's own sha256 (W5-ENG-MARGIN-V6's and this lane's observableRewrites entries both stand); check_twin_parity.py: OK. tests/twin_parity_inventory.json: W3B-ENG-GRID's rewrite of the same twin (test_dual_entry_placement_sizing_l4b, its MM_pair_defers row) met this lane's (its R2 MS-SF-A row) in one entry; the twin carries both hunks, the entry names both reasons, and its twinAssertionSha256 is the merged twin's (173 assertions). tests/CMakeLists.txt: git's clean merge put this lane's fixture if() inside W5-ENG-MARGIN-V6's block; each block is closed on its own. INT28 pick note: W6-ENG-FILL-ORDER 1 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ding market's quantity
The later of two opposite strategy.entry calls placed on one flat bar is one
transaction of its own quantity plus the earlier pending MARKET call's, and
TradingView costs that transaction at the signal close: past the equity it
drops the later call, however affordable its own quantity is. The engine's
explicit placement admission (entry(), src/source/pine_adapter.cpp:6019-6071)
priced the later call's own quantity only, so it kept calls TradingView
drops.
The admission now adds pending_opposite_market_units -- the own units of the
opposite market entries still pending on that bar -- for a flat
explicit-quantity call in the same_bar_market_tx_scope() family
(src/source/pine_adapter.cpp:6045-6055).
TradingView (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01;
tests/fixtures/same_point_entries): w6-f10g-pair-gross, one-contract MM-LF /
MM-SF / MS-LF / MS-SF pairs on 3000 of equity, seven passes: the later call
trades in exactly the 10 cells whose own-plus-pending transaction at the
signal close is within the equity (the other 18 book the earlier call alone).
w6-f10h-pair-gross-pyramiding1 is the same tape at pyramiding 1. The KI-65
probe on CME_MINI:NQ1! 15 (two contracts above ~25,000 cost more than its
equity) books the same rule: 145 later calls dropped, each transaction
costing more than the equity at its signal close -- or, for one buy that
fills first, at its fill, which the kernel's opening check already refuses.
Fail-before, the test built against 443e7605 (this lane's R1 tree): "FAIL
tests/test_same_point_entries_tapes.cpp:389 got == want" on both rows, "1549
passed, 2 failed". After: 1491 passed, 0 failed (its tape reading covers the
two new tapes).
test_dual_entry_placement_sizing's scope fence had pinned the NON_P2 mode
(pyramiding 1, a 55% wedge) as ungated: "a leaked gross gate would reject the
second". TradingView gates it (w6-f10h), so the fence now expects the later
call dropped; against 443e7605 it fails ("FAIL ... :674 non_p2.queued == 1 &&
non_p2.metadata_clean"). The observable rewrite's evidence and reason are
updated in tests/twin_parity_inventory.json.
Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No
kernel unit changes; no state added.
Population: Cloud Run experiment exp-w6-eng-fill-order-a-20260927 on this lane's
final engine tree, 2,166 probes (the 24 of families F10 and F12, 40 seeded
excellent controls and the static census of every rule's shape), against
exp-r5-tvdd-measure-20260927: pf-probe-ki65-dual-entry-precedence nq1-15
moderate -> excellent (engine 5091 trades -> TradingView's 4979).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 04368aa5e197223fcca90bf36293fb28781fd303)
INT28: every code, test and fixture hunk applies unchanged (each file's patch
is byte-identical to the lane's). scripts/check_twin_parity.py:
INVENTORY_SHA256 is the merged inventory's own sha256 (W5-ENG-MARGIN-V6's and
this lane's observableRewrites entries both stand); check_twin_parity.py: OK.
tests/twin_parity_inventory.json: the test_dual_entry_placement_sizing entry
carries this lane's reason and W3B-ENG-GRID's (its MM_pair_defers row), and
the merged twin's twinAssertionSha256 (173 assertions).
INT28 pick note: W6-ENG-FILL-ORDER 2
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… market pair fills first
At a process_orders_on_close close, with or without calc_on_order_fills,
TradingView fills a flat pair of opposite explicit-quantity MARKET entries in
the same order as at an open: the buy first, as one transaction of its own
quantity plus the pending sell's, then the sell of its own quantity. The
tape's two long rows both carry the buy's signal (TAIL-TRIAGE
F10-KI65-DUAL-ENTRY, the COOF+POOC opposite-market probes).
The adapter's close pass for such a pair,
apply_terminal_explicit_market_policy (src/source/pine_adapter.cpp:15986),
re-executes the candidates in source order with every later opposite call a
full ReverseTo, so a sell-then-buy pair booked a zero-length short closed by
the buy. For exactly that pair from flat the pass now executes the buy first
as Transact(own + pending sell), then the sell as Transact(-own)
(src/source/pine_adapter.cpp:16094-16119). The net position is unchanged; a
buy-then-sell pair keeps the source order, event for event.
TradingView (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01;
tests/fixtures/same_point_entries): w6-f10d-pooc-pair and
w6-f10e-pooc-coof-pair are one file; their market cells (MM-LF, MM-SF,
MM-SF-Q) are replayed. Their priced cells and w6-f10f-coof-pair
(calc_on_order_fills alone) are read off the tape only: the adapter's POOC
close pass neither orders a priced leg against the market nor adds the
pending market's quantity there, and outside same_bar_market_tx_scope() the
open pair is not ordered (README).
Fail-before, the test built against 04368aa5 (this lane's R1+R1b tree):
"FAIL tests/test_same_point_entries_tapes.cpp:415 got == want" on both rows,
"2106 passed, 2 failed". After: 2108 passed.
test_pooc_coof_reversal_gross_admission pinned the scratch row's side as the
first call's ("scratch.is_long == first_long", while its header said the side
is only report attribution and not asserted); TradingView's scratch row is
the buy's either way, so it asserts scratch.is_long. Against 04368aa5 it fails
("FAIL ...:160 scratch.is_long", 146 passed, 2 failed). A frozen twin
assertion: tests/twin_parity_inventory.json records the observable rewrite.
Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No
kernel unit changes; no state added.
Population: Cloud Run experiment exp-w6-eng-fill-order-a-20260927 on this lane's
final engine tree, 2,166 probes (the 24 of families F10 and F12, 40 seeded
excellent controls and the static census of every rule's shape), against
exp-r5-tvdd-measure-20260927: pf-probe-coof-pooc-opposite-market-ordering on
es1-15, eurusd-15, nq1-15 and xauusd-15 and
pf-probe-coof-pooc-opposite-market-fran-factors on eurusd-15 and xauusd-15
moderate -> excellent.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
(cherry picked from commit cdbeb0b0e80238617f88cce087c95730ca9bec1b)
INT28: every code, test and fixture hunk applies unchanged (each file's patch
is byte-identical to the lane's). tests/twin_parity_inventory.json: this
lane's test_pooc_coof_reversal_gross_admission entry and W5-ENG-MARGIN-V6's
test_pooc_open_money_event entry met at one place; both stand, in key order.
scripts/check_twin_parity.py: INVENTORY_SHA256 is the merged inventory's own
sha256; check_twin_parity.py: OK.
INT28 pick note: W6-ENG-FILL-ORDER 3
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…n a default reversal is declined On one bar a held position's protective strategy.exit stop and a reversing strategy.entry stop on the same side of the price fill in the order the bar's path reaches them: the reversal at its level first when its level comes first (and the exit is then void), the exit first otherwise, the magnifier changing nothing. When both are already through at one opening TradingView fills the exit first there, whichever was placed first, and the entry then trades its own frozen quantity from flat (TAIL-TRIAGE F12-INTRABAR-STOP-PRIORITY). The engine books every one of those cells at a fixed size. Under the v6 default size a reversal can be declined at that opening, and defer_declined_reversal_exits_at_adverse (src/source/pine_adapter.cpp:14803) cancelled every full-position exit stop the bar's adverse extreme reaches, or re-issued it at that extreme -- a stop the open was already through included. TradingView fills such a stop at the open. It is no longer deferred (src/source/pine_adapter.cpp:14887-14893); a stop the open has not reached keeps the legacy deferral. TradingView (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01; tests/fixtures/same_point_entries): - w6-f12a-stop-priority (fixed 1): A1/A2 (a falling bar, open nearer the high / the low), A3 (exit above the entry), A4/A5 (both through, entry / exit placed first), A6/A7 (only one through), B1/B2 (short mirrors). The engine already booked all of them. - w6-f12c-stop-priority-magnifier: the same trades with use_bar_magnifier. - w6-f12b-stop-priority-default (v6 default): in B2 the short, margin-called in part on its entry bar, declines its reversal at the open, and the protective stop fills there (1477.89); the engine had cancelled it and held the short to the cleanup. A4 stays out of the row: after the exit TradingView opens the reversal's frozen 2x short and margin-calls it at the same price, which the engine's margin model refuses (README). Fail-before, the test built against cdbeb0b0 (this lane's F10 tree): "FAIL tests/test_same_point_entries_tapes.cpp:492 got == want" on the w6-f12b row, "2386 passed, 1 failed"; the fixed-size and magnifier rows pass before and after. After: 2387 passed. test_adapter_live_state_equivalence: 4 of its 108 transcript digests move, re-harvested with PINEFORGE_V19E_HARVEST (the base 8633d944 reproduces every old value) -- exactly the four "reversals" configurations in which this rule acts, counted by an instrumented copy of the adapter over the battery; listed in tests/test_adapter_live_state_equivalence_data.hpp. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No kernel unit changes; no state added. Population: Cloud Run experiment exp-w6-eng-fill-order-a-20260927 on this lane's final engine tree, 2,166 probes (the 24 of families F10 and F12, 40 seeded excellent controls and the static census of every rule's shape), against exp-r5-tvdd-measure-20260927: no F12 probe's trades move (10 excellent; the 4 strong ones first diverge at TradingView margin calls, family F09). 27 census probes change only their 'Engine entry incarnation' column, by -1 per firing: the stop the rule leaves in place is one request the old deferral re-issued. Their trades and grades are identical. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit d46096ecabc655f0b64e15368142b8bf1f4952ba) INT28: the rule, the tape rows and the fixture README apply unchanged (each file's patch is byte-identical to the lane's). Its re-pin of four tests/test_adapter_live_state_equivalence_data.hpp digests (harvested on the lane's tree) is not applied here: INT28 re-harvests every behaviour pin once, on the integrated tree, and attributes each moved row to its rule. INT28 pick note: W6-ENG-FILL-ORDER 4 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…r as one transaction, admitted without its commission TradingView books a flat opposite pair with a commission exactly as W6-ENG-FILL-ORDER's zero-cost tapes: the same order at one fill point, the later call as one transaction of its own quantity plus the pending market's, that order's fee split across its rows by quantity (w6-f10i, w6-f10j, w6b-p1d). The transaction is admitted against the equity at the signal close with the commission left out: every pair under the equity by less than its fee trades both legs, every pair just over it drops the later call, with and without a commission (w6b-p1a/p1b/p1c). A pair whose first fill is the whole transaction is booked with no margin call although the fee leaves the equity short of that fill: TradingView checks the margin once both legs have filled. The pair rules now hold on same_point_pair_scope(), the configurations of the zero-cost batch scope with a commission; same_bar_market_tx_scope() is that scope with zero cost, so the batch itself is unchanged. Outside the batch the later market call is one Transact of both quantities, and the opening margin checkpoint of a pair's first fill waits for the opposite leg filling at the same price (same_point_pair_fill_follows). test_same_point_entries_tapes replays the six tapes (4217 passed); against b3132a52 the five commissioned rows fail at :603 (got == want) and the no-commission control passes. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 2fc453375636132da3be6db5200885f9a45afa7d) INT28: every code, header, test and fixture hunk applies unchanged (each file's patch is byte-identical to the lane's), beside W5-ENG-MARGIN-V6's margin code in on_applied's opening checkpoint, which git merged without a conflict. The two are scope-disjoint at that fill: same_point_pair_scope() excludes process_orders_on_close and slippage, which W5's C2 and M2, and MK, require; both lanes' tapes pass. INT28 pick note: W6B-ENG-PAIRS 1 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… and a flat pair books as under 0 At one fill point TradingView fills lots of one side by rank -- market, stop, limit -- under pyramiding 2 (w6-f10c) and 3 (w6b-p2c: three legs in all six placement orders), and books both W6 pair tapes under pyramiding 2 byte for byte as under 0 (w6b-p2a, w6b-p2b). same_point_pair_scope() now covers any pyramiding (the zero-cost batch scope keeps pyramiding <= 1), so order_same_point_entries ranks same-side lots and the pair rules hold. The legacy pyramiding-2 route for a book that is not an exact market pair re-sized a pair's market leg into a full reversal at the fill (SM-LF: the sell reversed the buy stop's long); for the exact pair book -- one market call beside one priced opposite call of its bar, nothing replaced, cancelled or resting from an earlier bar -- flush_pending_same_bar_commands keeps the market leg's frozen transaction, told by its priced opposite caller when that one follows. Every other pyramiding-2 book keeps the legacy source order and batch-time admission the MM cases of test_dual_entry_placement_sizing_l4b pin. test_same_point_entries_tapes adds four rows (5744 passed); against f4148f9d exactly those four fail at :650 (got == want). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 63fabaa19ab9b17dbd1f3c43185ee69bff038017) INT28 pick note: W6B-ENG-PAIRS 2 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…inst the market leg in TradingView's order At a process_orders_on_close close TradingView fills a flat pair of one market call and one priced opposite call the close has reached in W6's rank order, the later call one transaction of its own quantity plus the earlier call's when that one is the market (w6-f10d-pooc-pair and w6-f10e-pooc-coof-pair: MS-SF buys 2 with the stop first, ML-SF sells the market first and buys 2 with the limit, SM-SF / LM-SF buy the market before the sell stop or limit, LM-LF sells the market before the buy limit). The close pass filled the priced leg first (fill_pooc_close_entries) and the market leg after it, each with its own quantity. fill_pooc_close_pair now fills exactly that pair -- pyramiding 0, fixed default, explicit quantities, nothing else entry-like on the book -- at the close in rank order before the rest of the pass; a market pair keeps apply_terminal_explicit_market_policy. test_same_point_entries_tapes now covers every cell of both tapes (5768 passed); against 07c87619 the two rows fail at :650 (got == want). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 287a1ffee358f06444485c93152a7627ad363d9d) INT28 pick note: W6B-ENG-PAIRS 3 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ngView's order as one transaction With calc_on_order_fills and without process_orders_on_close TradingView books a flat pair placed once per bar exactly as without it: W6's two open-pair tapes re-exported under calc_on_order_fills are byte-identical (w6b-p4a, w6b-p4b), and so is the close-pair tape's next-open counterpart (w6-f10f). same_point_pair_scope() now covers calc_on_order_fills for calls placed outside a fill recalculation (the zero-cost batch scope keeps excluding it), so the later market call trades its own quantity plus the pending market's and order_same_point_entries fills a buy placed after a sell first. The engine had booked MM-LF, MM-SF, MS-SF, SM-LF, LL-SF and MS-SF-Q in placement order with own quantities. test_same_point_entries_tapes adds three rows (6582 passed); against 57c17300 w6-f10f and w6b-p4a fail at :660 (got == want), and w6b-p4b, which the engine already booked, passes. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 50152b883d396e3d9d17e4a833cdf24c06e245cc) INT28 pick note: W6B-ENG-PAIRS 4 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ed bar at or before the calling bar's time
request.security(<tf finer than the chart>, expr, lookahead_on) merges the
requested series at the calling chart bar's time: TradingView reads the
requested bar that opens at or before that instant, the latest such bar.
Where a requested bar opens at the chart bar's time that is the first
intrabar (the round-7 rule, test_ltf_lookahead_first_bucket, BTCUSDT 1D).
OANDA stamps its XAUUSD daily bars at 17:00 ET, inside the 17:00-18:00
break, an hour before the session each bar carries: no requested bar opens
at the stamp, and TradingView reads the previous chart bar's last one.
TradingView, on synthetic read-out probes written for this lane (lab tv
--no-note, 1D, 2025-04-01 .. 2026-05-01; tests/fixtures/ltf_lookahead_calling_open):
- w8c-ltfon-xau1d (OANDA:XAUUSD): "15" lookahead_on merges the 16:45 ET bar
on 211 bars (-15 min), Friday's 16:45 on the 46+4 Sunday stamps (-2895),
and the last bar of an early close or long weekend otherwise; gaps_on +
lookahead_on reads the same bar; lookahead_off the chart bar's last.
"The latest 15m bar opening at or before the stamp", on the xauusd-15
lane's own TradingView bars, reproduces all 278 read-outs.
- w8c-ltfon-tfs-xau1d: "60" reads 16:00 ET (-60), "5" 16:55 ET (-5); "240"
reads 0 on every bar, because TradingView's 4h grid opens at the 17:00 ET
stamp itself (lookahead_off: 13:00 ET).
- controls w8c-ltfon-eth1d, -eur1d, -aapl1d: a 15m bar opens at the stamp
(00:00 UTC, 17:00 ET, 09:30 ET), and every bar reads it (0).
- w8c-ltfon-3m-xau1d ("3"): 16:57 ET on 215 bars; on the Black Friday bar it
reads Thanksgiving's 21:57Z bucket, which holds the 21:59 minute alone and no
completion rule closes, and on 2025-07-08 the 20:57Z bucket of two minutes: a
pending last bucket is still the bar at or before the stamp.
The engine's first-bucket latch (calling_open_latches_first) always fed the
slice until its first bucket was published, so on the XAUUSD daily stamp it
read the 18:00 ET bucket: one requested bar late ("15": +60 where TradingView
reads -15; "60": +60 for -60). feed_aux_security_for_chart_bar now closes the
latch before feeding an auxiliary bar whose requested bucket opens after the
calling bar's time, so the chart body reads the previous slice's last bucket
and the slice follows the body. The same closing applies whenever the next
auxiliary bar opens a bucket after the calling bar's time while the latch is
still open, and a bucket no completion rule has closed yet -- the previous
slice's last (the Thanksgiving singleton) or this slice's first, when its last
minute did not print -- is that bar: it is finalized and published first.
lookahead_on already evaluated it on its sub-bars, so its completion rewrites
that slot (sub-bar count at least 2, as a boundary emission does). Before,
such a bucket was emitted only on the next bucket's first bar, whose partial
evaluation then reached the chart body. Heikin-Ashi sites keep the
established path. The host header's comment says the rule in the same two
lines.
Moved pin: tests/test_split_feed_partial_bucket.cpp case (e) asserted that
the lookahead_on "3" twin reads each XAUUSD daily bar's own first bucket; its
header says the tape behind it pinned lookahead_off only. It now asserts
TradingView's reads from w8c-ltfon-3m-xau1d: nothing on the range's first
bar, the 16:57 ET bucket (close 4163.575) on Thanksgiving, the finalized
21:57Z singleton (close 4158.8) on Black Friday. Its lookahead_off cases are
unchanged. The test is an A24 row (scripts/check_twin_parity.py A24_NAMES),
outside the twin-parity freeze.
Fail-before: tests/test_ltf_lookahead_calling_open_tapes.cpp built against
8633d944's library: "44 passed, 10 failed" (4 XAUUSD "15" bars, 4 XAUUSD
"60" bars, the synthetic 1m -> "5" aggregation twice); every control bar
passes. After: "54 passed, 0 failed". test_split_feed_partial_bucket's case
(e) fails on 8633d944 ("dense.on_count == 0") and, without the tail
finalization, on Black Friday ("friday.on_count == 459 + 438"); it passes. Engine read-outs of the same probes on
the xauusd-1d lane feeds (1D chart + the 1m auxiliary slice): 277/278 equal to
the tape for both XAUUSD probes (the one left is the first bar, where no
auxiliary bar precedes the chart), the controls unchanged.
Population, Cloud Run exp-w8c-security-fix2-20260928 (this tree) against
exp-r5-tvdd-measure-20260927, 139 probes: security-rsi-cross-threshold
(xauusd-1d, "15" ta.rsi(close, 14)[1] with lookahead_on) weak 33.3% (engine
4 / TV 6 trades) -> excellent 100% (6 / 6);
3commas-3commas-triple-rsi-dca-long-risk-adjusted-averaging-strategy and
triple-rsi-dca-trace (eurusd-1d, "5" / "15" / "30" ta.rsi(close, 7)[1]
with lookahead_on: on 36 days the first 15m bucket, and on 24 the first 30m
bucket, loses its last minute after the 17:00 ET rollover and stays pending
until the next bucket) strong 83.3% -> excellent 100%. The 14 other W8C targets,
40 seeded random excellent probes of the same lanes and the other 82
population probes with a finer lookahead_on site are byte-identical
(trade-set sha).
Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2). No
kernel unit changes and no durable state is added (first_bucket_published
already exists and is hashed).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 1fa82e94d6794a3d8a94bf2cb3e5c7179e517001)
INT28: every code, header, test and fixture hunk applies unchanged (each
file's patch is byte-identical to the lane's). tests/CMakeLists.txt: git's
clean merge put this lane's fixture if() inside W6-ENG-FILL-ORDER's block;
each block is closed on its own.
INT28 pick note: W8C-SECURITY 1
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ainst the closed position Under process_orders_on_close, a strategy.close_all() or a whole-position strategy.close(id) the script issues in the same pass as a strategy.entry settles first. A same-side entry called after the close is judged against the position the close flattens: its margin is the resulting position, (held + own) x close x pointvalue x margin %, and TradingView drops it when that exceeds the equity. An opposite-side entry is not charged for the held lots. An admitted entry opens after the close as a fresh position, so pyramiding=0 does not block it; and a same-side entry the script called before the close survives it too, when the cap admitted it at its call. The adapter charged the entry for its own quantity only (an ordinary POOC close_all fills at its call, so the affordability gate saw a flat book). After a strategy.close(id) it dropped entries TradingView fills: under pyramiding=0 as a same-side add at the call, and in every case when the queued close flattened the position -- flush_pending_closes cancelled the pass's same-side market entries whichever call came first (ab9714b kept that blanket cancellation for flush-time close(id) while preserving under-cap entries for a POOC close_all). entry() now charges a same-side entry for the position before the pass's closes (the script's position view, which the close_all froze when it filled) without changing the recorded projection, and waives the pyramiding=0 same-side no-op when the pass's queued closes flatten the position; flush_pending_closes no longer cancels the pass's entries. Scraped pf-probe-concord-frangestate (CME_MINI:NQ1! 15, strong) closes its position with close_all and re-enters on the same side in one pass on 22 bars: TradingView drops the entry on the 15 bars where (held + qty) x close x 20 exceeds the equity; the engine re-entered on all 22 (its 15 engine-only trades). Tapes, tests/fixtures/pooc_same_pass_close (lab tv --no-note, BINANCE:ETHUSDT.P 15, the corpus 15m bars of 2025-04-01 .. 04-09 embedded): - w8a-pooc-sameside-closeall: A (close_all + long 50 %), D (close("H") + long 50 %) and E (the short mirror) are dropped; B (30 %) and D2 (close("H") + 30 % under pyramiding=0) fill after the close; C (opposite) is not charged; V pins the held lots at the close price. The whole year replays 589/589 (lane scratch). - w8a-pooc-close-reentry-pyr2 (pyramiding=2): close("L") then entry("L_add") on even days, the other order on odd days; L_add opens at the close and is held to the cleanup on every day. Pinned by tests/test_pooc_same_pass_close_tape.cpp: 1906 passed, 2 failed on 8633d944; 1926 passed, 0 failed here. The legacy integration row test_strategy_close_pooc_cancels_same_bar_ market_reentry (close("L") then entry("L_add"), pyramiding=2) pinned the blanket cancellation; it is the second tape's close-first shape, so it becomes test_strategy_close_pooc_keeps_same_bar_market_reentry and expects the held L_add (position 1). Its frozen twin assertion is an observable rewrite citing the tape (tests/twin_parity_inventory.json, INVENTORY_SHA256 re-pinned in scripts/check_twin_parity.py; the legacy base copy is edited alike): check_twin_parity OK, 380 pinned observable rewrites. Pine policy, adapter only: no kernel change, no ADR-0001 ruling. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit c19ad463408aabd01510acb8123cab3b8e276e3b) INT28 (with W3B-ENG-GRID picked before it): tests/twin_parity_inventory.json: W3B-ENG-GRID's rewrite of the same twin (test_integration_l4d, its close-reservation rows) met this rule's (its test_strategy_close_pooc_keeps_same_bar_market_reentry row) in one entry; the twin carries both hunks, the entry names both reasons, and its evidence is the merged twin's (569 assertions); check_twin_parity.py: OK. INT28 pick note: W8A-SIGSTATE-1 3 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…every call TradingView checks a default-quantity LIMIT strategy.entry that opens from flat at every call -- its placement and each re-issue -- against the calling bar's close: a call whose quantity costs more than the equity there is rejected, and it takes the resting order of that id with it, however affordable the order's own level is. A call it admitted is not re-checked on later bars, and the fill-time check still refuses a fill whose own notional exceeds the equity. The adapter left LIMIT entries out of its placement check (affordability_scope), so a re-issue TradingView rejected kept its resting order alive. entry() now prices a flat pure-LIMIT entry under a default fixed quantity at the calling close, and drops the call and cancels the resting same-id request when it is unaffordable, as it already did for a pure STOP entry. The order's own affordability flags are unchanged. Scraped algotorma-algo-torma-orb-strategy (BINANCE:BTCUSDT 15, strong) re-issues its breakout limit every bar while armed; its last re-issue (2026-01-14 23:45, close 96951.78 over an equity of 96846.71) was rejected in TradingView, and the engine filled the resting order on 01-15 03:15, its one engine-only trade. Tape: tests/fixtures/limit_entry_affordability/w8a-btc15-limafford (lab tv --no-note, BINANCE:BTCUSDT 15, equity 95000, default fixed 1; the btcusdt-15 lane's bars of 2025-04-25 .. 04-30 embedded). A (re-issued, last calls unaffordable) never fills; C (placed once affordable, later closes above the equity) fills; E (level above the equity) never fills; B is the control. Pinned by tests/test_limit_entry_affordability_tape.cpp: 27 passed, 1 failed on 8633d944; 25 passed, 0 failed here. Pine policy, adapter only: no kernel change, no ADR-0001 ruling. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 62bc9cb0bb388b3dcf7e65d1d0559e8028617801) INT28: every code, test and fixture hunk applies unchanged (each file's patch is byte-identical to the lane's). src/source/pine_adapter.cpp, entry(): this rule's block and W5-ENG-MARGIN-V6 M1's placement check were inserted at the same place; both stand, M1's first. Their scopes are disjoint -- M1 judges a default percent_of_equity MARKET add to a held position of its side, R3 a default FIXED LIMIT entry from flat -- so no call reaches both. tests/CMakeLists.txt: each fixture block is closed on its own. INT28 pick note: W8A-SIGSTATE-1 4 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…es only its own quantity A strategy.entry called again with the same id on the same bar replaces the order its first call placed. When that first order is a MARKET order that reverses the position, TradingView sizes the reissue against the position the first order would leave: the reissue is a plain transaction of its own quantity X against the held side P -- it closes up to X of P and opens only the excess -- not a reversal. It does so on either side, for every quantity kind (default percent, fixed and cash defaults, an explicit qty), with and without process_orders_on_close. On the buy side the held short keeps its bracket; on the sell side a default-size reissue voids the held long's bracket. A single call, a LIMIT first call re-issued as MARKET and a strategy.cancel between the calls keep the full reversal. Lanes W8A-SIGSTATE-1 (7d668c80) and W8B-SIGSTATE-2 (aa9ee275 + d339b4de) wrote this rule independently; each failed the other's tapes (W8A's missed the fixed-default sell that voids the held long's bracket, W8B's the explicit qty and cash reissues). This is the one rule both tape sets pin: - entry(): a same-id MARKET reissue of a reversal whose predecessor is still staged in the callback's same-bar batch (every fixed default, and a variable size over a single-lot short) records the replacement (replaced_opening, replacement_predecessor_market), so a default-size sell voids the held long's brackets as the live-predecessor path already did; its staged transaction carries its own units only, and the batch row records the held units it left out (PendingSameBarCommand::staged_reversal, staged_reversal_held_units, folded into the source state hash only when set) -- W8B's mechanism, widened from percent/fixed defaults to every quantity kind the batch admits; - flush_pending_same_bar_commands(): the round-8 short-seed book (short seed, then Long, Short, close(Long), close(Short) on one bar) keeps its full reversal -- the flush puts the held units back and clears the replacement facts, W8B's d339b4de scope (tests/oracle/test_oracle_short_seed.cpp, GateControl::SameIdReplacement); - resolve_terms(): replacement_transaction covers every MARKET reissue of a live MARKET predecessor, on either side and for every quantity kind -- W8A's mechanism; a priced reissue keeps the old sell default-percent scope. Layer: the source-adapter parity runtime (ADR-0001, Decision, layer 2); no kernel unit changes. Tapes: the fifteen lab tv --no-note exports of both lanes (BINANCE:ETHUSDT.P 15), byte-identical copies, in tests/fixtures/same_bar_reissue, replayed by one test, tests/test_same_bar_reissue_tapes.cpp. Built against 8633d944's library: "679 passed, 10 failed" (5 of 15 tapes: the four single-call controls and w8b-flip-double-sell); on this commit: "689 passed, 0 failed" (15 of 15). test_replaced_percent_short_market and its oracle twin pinned an explicit-qty reissue as a full reversal to SHORT 2 with no tape behind it; the row now expects TradingView's 3 - 2 = 1 LONG (tapes w8a-dbl-p10 cell C, w8a-dbl-scope3 cells G and H; same CHECK count), and tests/oracle.sha256 re-pins that file's digest and the tree digest. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 93c840f58b5232f56d8ce969a4b750c1885a1292) INT28 pick note: R1-CONSOLIDATE 1 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…e the buy LIMIT entries it reached At one opening price TradingView fills every MARKET order first, entries and exits alike (a strategy.close_all queued at the previous close among them), then the buy LIMIT entries the open has already reached, lowest limit first -- whether a limit was placed at the previous close or has rested since an earlier bar; neither the order they were placed in nor their ids decide. A limit filled that way keeps the quantity it was placed with. The native matcher breaks a same-point tie by queue order, so the engine filled them in placement order: a limit resting beside a close opened and was closed by it at once, and limits placed in descending order filled in that order. This is lane W8A-SIGSTATE-1's pinned patch R-A. Measured on a scratch tree that also carries W6 (b3132a52), W6B (25ed4413) and W8A's R-B (6c5de611, which R-A's patch was written on): without R-A the engine books w8a-dca-open E1 as B, A, C (TradingView: B, C, A), closes E2a's L1 with the close_all at once (TradingView holds it) and fills E4b's HL before HM; W6's same-rank tie-break is placement order and W6B's rules leave these books alone. With it, W6/W6B's tape test passes 6582/0 and W8A's full w8a-dca-open test 140/0. on_bar_open() now re-prices every live pure-limit buy entry the open has reached in place (order_open_marketable_limit_entries, over the kernel's existing ReplaceOptions::keep_handle), lowest limit first, whenever a MARKET request stands behind one of them or they are not already in ascending limit order; a re-priced entry books the opening print (forced_execution_price). Not under process_orders_on_close, calc_on_order_fills, streams or a fill recalculation; sell limits and stop entries keep the queue order (no tape covers them). A flat book holding MARKET entries of both sides is left to the opposite-pair route (lanes W6/W6B): TradingView fills that pair before the limit too but also books the pair's later call as one transaction admitted at its gross (tape r1c-ra-prior-limit cells M and N), which is not this rule; the frozen legacy pin of that shape (test_dual_entry_placement_sizing's MM prior-bar gapped limit) keeps its route. Tapes (tests/fixtures/open_fill_order): w8a-dca-open (a byte-identical copy of W8A's pyramiding_open_order tape) and r1c-ra-prior-limit, a new lab tv --no-note export (NYSE:F 1D, 2024-12-01 .. 2025-08-01): buy limits resting since d0 inside the gap between d1's low and d2's open, market orders placed at d1's close. tests/test_open_fill_order_tapes.cpp replays cells E1, E2a, E2b and E, X, O trade for trade. Built against cd634e62's library: "213 passed, 5 failed" (E2b, the control, passes); on this commit "218 passed, 0 failed". The R-B cells (E3, E4a, E4b) and the pair cells (M, N, P) are read off TradingView's rows only. Pine policy in the adapter over the kernel's existing keep_handle re-price: no kernel change, no ADR-0001 ruling. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 2239825045e2326b5b27ccf5ce3b53a4178650f1) INT28 pick note: R1-CONSOLIDATE 2 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ADR-0001's rows for ReplaceOptions::keep_handle (R5 lane V19-D's replace row, and K-OCA-KEEP's group-effect receipt order) said the Pine adapter never sets keep_handle. R-A's opening-price order re-prices the buy limits an opening price reached with it (order_open_marketable_limit_entries), so the V19-D row now names that use, the K-OCA-KEEP row says the adapter reaches the group-effect order only when such a limit is a group member, and K-OCA-KEEP's own "corpus parity unchanged by construction" reads "the Pine adapter then never setting keep_handle". Wording only; no anchor on these cells. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 5ce0dd28e2aeb0cb33781a464d288d7813e6d679) INT28 pick note: R1-CONSOLIDATE 3 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…es from it TradingView's deep backtest (ws report channel) computes the script from the first chart bar at or after the requested range start and from no earlier bar -- metrics.json records it as wsProvenance.returnedRange.from -- and its broker is live from there: an order placed on any bar before the tape's first fill can make that fill. run_strategy.py gated every tape-graded run at the bar before TradingView's first entry, so such an order was never placed (jaysharma x2: a supertrend stop placed before 08:00 UTC fills at 08:15; ki62 eurusd-15/xauusd-15: phase A's stop, armed once). Pinned by 13 lab tv --no-note tapes of synthetic probes (tests/fixtures/run_harness_window): bar_index 0 is the returned range start on ETHUSDT.P / EURUSD / XAUUSD / ES1! 15 (00:00 UTC), NYSE:F 15 (13:30), NSE:NIFTY 15 (03:45), BTCUSDT 1D (00:00), XAUUSD 1D (21:00); a stop, a limit and two market entries placed on 2025-03-31 bars never trade; a stop placed at 01:00 UTC makes the 06:30 first fill. The rule (_tv_first_bar, _tv_entry_emit_window, main): a run whose loaded feed starts on TradingView's first bar opens its window there; unless it enters before the bar the tape reports from (TradingView's first entry less one bar) -- TradingView filled nothing before its first entry, so its state is not TradingView's -- when it runs again from the signal bar (_entered_before_tv); a re-run whose signal-bar gate would still admit the feed's first bar (the engine admits one bar before a gate) would repeat the run and is skipped. A warmed, padded or later-starting feed keeps the signal-bar start. Reported rows are unchanged. Measured on Cloud Run over the whole population f30f1ab8 (7994 probes, codegen d7e095f): 141 trade sets move, 10 up (jaysharma x2 on data, NIFTY-15 and EURUSD-1D; ki62 eurusd-15, xauusd-15, btcusdt-15; job-2738), 0 down, 0 probes with any worse canonical metric (exp-run-harness-win-nf-20260928 vs exp-run-harness-before-20260928). Without the entered-before fallback the same rule drops one probe (exp-run-harness-win-b); opening the gate on every run holding TV's first bar drops nine (exp-run-harness-win-a). tests/test_run_harness_window_tapes.cpp replays the ETH tapes through the Pine adapter; scripts/test_run_strategy_emit_window.py pins the harness decision and reads the rule off every tape (fails on the base: 1 failure, 7 errors). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit b0976fa7117c8179ed8a13eec7df0b73cdc94df1) INT28 pick note: RUN-HARNESS 1 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…magnified TradingView backtests a script that declares strategy(..., use_bar_magnifier = true) on its bar magnifier: its tape of mi-fx-eth-15 and of the same script declared off (w9mag-fx-eth-15-off, lane CG-W9-MISC, copied here) enter alike and part on 7 of the 23 replayed exits. Codegen e8f6648 (cg/w9-misc) makes a declaring library export int strategy_declares_bar_magnifier(void); this is the host half of CG-W9-MISC's proposal. - run_strategy.py reads the export (Strategy.declares_bar_magnifier). On an intraday chart coarser than 1m, with the case runner's PINEFORGE_RUN_MAGNIFIER_FEED (held to _SHA256 when set) and runtime_overrides.bar_magnifier not false, the engine's input becomes that 1m feed with input_tf=1, script_tf=<chart>, the magnifier on, read over exactly the chart feed's loaded bars (from the first chart bar's open to the last minute of the last one, so the range-end bar is whole and an earlier 1m start warms nothing the chart run does not); TradingView's window is still walked on the chart feed, and the trace JSON names the feed the engine read. Daily, weekly and monthly charts log `declared-not-run` (TradingView keeps its own daily bars as chart bars; no engine input yet pairs them with a 1m intrabar source), as do a run with an auxiliary request.security feed and a missing feed. --runner docker never reads the export and says so when the environment names a magnifier feed. - pineforge.h declares the export in the codegen-emitted group; check_c_abi_runtime.py: 67 public declarations (was 66), 58 runtime implementations; README and the ABI pages count it. Two count sentences are rewritten, not dropped: docs/coverage.md's "Every compiled PineForge strategy `.so` exports that public set." (e7cdf05) now names the one export only a declaring script has, and docs/pages/index.md's "`<pineforge/pineforge.h>` -- **66**: 58 runtime implementations plus eight per-strategy generated exports." (3cb0a18, INT23) counts 67 and nine. - tests: scripts/test_run_strategy_magnifier.py (fails on the base at import); test_adapter_magnifier_intrabar_tapes section 4 replays the twin on the 15m chart bars (23 of 23 exits the twin's). Measured on Cloud Run with codegen e8f6648 and a measurement-only copy that also reads the verifier's PINEFORGE_VERIFY_FEED_1M (the file the runner will name; the deployed runner does not set the new variable yet): bystry1991 btcusdt-15 strong->excellent, es1-15 weak->moderate, xauusd-15 moderate->strong, eth-scraped-15 (hard) moderate->strong; the five corpus magnifier probes stay excellent (price-exact 99.5->100); the daily lanes are unchanged; 18 controls byte-identical; 0 down (exp-run-harness-mag-item2-20260928 vs exp-run-harness-mag-base-20260928). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 2c397248fbb9d107bc3a75c4f9e12677db1d1de5) INT28: main 962960b already carries lane XSYM-D's four runtime setters (62 runtime implementations, 70 public declarations of pineforge.h, 113 exports with the native host's 43); this lane adds the ninth generated export, so the integrated counts are 71 declarations (62 runtime implementations plus nine generated exports) and 114 with the magnifier export. scripts/check_c_abi_runtime.py (EXPECTED_PUBLIC_DECLARATIONS 71, EXPECTED_RUNTIME_IMPLEMENTATIONS 62), scripts/test_ci_verify.py (both counts, XSYM-D's setter checks and this lane's generated-export assertions), README.md, docs/coverage.md, docs/pages/abi-stability.md and docs/pages/index.md carry both lanes' text with those counts; check_c_abi_runtime.py: exit 0. scripts/test_run_strategy_requests.py (lane XSYM-D, on main): its _FakeStrategy gains declares_bar_magnifier = False, as this lane gave the fakes of its own base, since run_strategy.main() now reads the flag (15 of 15 pass). INT28 pick note: RUN-HARNESS 2 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ty step
TradingView reports syminfo.mincontract equal to the symbol's quantity step
on every campaign lane: 18 lab tv --no-note read-outs of one synthetic
script (lane W2-CG-LOWERING-TRIO, finding F06; tests/fixtures/
syminfo_mincontract) -- BINANCE:ETHUSDT.P 0.0001, BINANCE:BTCUSDT 0.00001,
OANDA:EURUSD / XAUUSD 0.01, NASDAQ:AAPL, NYSE:F, NSE:NIFTY, CME_MINI:ES1!,
NQ1! 1, on 15 and 1D. Codegen with F06 reads it through
get_syminfo_metadata("mincontract").
inputs_run_kwargs declares syminfo_metadata["mincontract"] from the lane
template's step, PINEFORGE_VERIFY_QTY_STEP in the environment the case
runner builds for every case (lane_input_templates.environment), when it is
a positive finite number -- a probe that declares its own keeps it.
--runner docker passes no syminfo_metadata (a documented gap) and now says
so. Only a declared step counts: the two ETH
15m templates declare none, and the Lab's 0.0001 there is its own default,
which its provenance records as not lane evidence; with it, the hard probe
data/blitz-locked-macd-pullback-sniper-trend-adx-filtered-strategy drops
strong->moderate (exp-run-harness-mc-item3-20260928), so the rule is
narrowed to templates that name their step.
Measured on Cloud Run on the 12 blitz probes (the population's only
mincontract readers, census over 1384 sources) and 12 excellent controls,
codegen d7e095f + W2's F06 lowering: btcusdt-1d, btcusdt-15, ethusdtp-1d,
xauusd-1d minimal->excellent, xauusd-15 strong->excellent, 0 down,
controls byte-identical (exp-run-harness-mc-item3b-20260928 vs
exp-run-harness-mc-base-20260928). With codegen d7e095f alone
(mincontract lowered to na) nothing moves.
scripts/test_run_strategy_mincontract.py reads every read-out (fails on the
base at import).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 966df9b50eb695761de4301d087682941eb4c19c)
INT28 pick note: RUN-HARNESS 3
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… composition Git merged every wave-J lane into src/source/pine_adapter.cpp without a textual conflict, but several lanes changed the same functions: - on_applied: TVDEF-DROPS R3, W3-ENG-EXIT-ALLOC rules B and C, W5-ENG-MARGIN-V6 M2, C2 and MK, W8E-EXITS C1; - the exit-reservation functions (W3 rule A) beside W5's MR and W8E C1; - entry: W5 M1, W6-ENG-FILL-ORDER F10 (1 and 2) and W8D-NOTRADES. Their scopes exclude one another -- M1 needs a percent_of_equity default and a held position of the add's side outside process_orders_on_close and calc_on_order_fills; F10 needs a flat book inside same_bar_market_tx_scope(), which admits a percent default only beside a held short seed and never calc_on_order_fills; W8D needs a calc_on_order_fills recalculation -- so no script reaches two of them. - defer_declined_reversal_exits_at_adverse: TVDEF-DROPS R2 and W6 F12, which W6 already built on. - on_bar_open: R1-CONSOLIDATE's R-A (the buy limits an opening price reached fill after the market orders, lowest first) and the entries W8A-SIGSTATE-1's R-B admits (a pending limit does not count against pyramiding). test_int28_rule_compositions replays four TradingView tapes through the Pine host and requires every trade each tape closes to be the engine's -- three synthetic probes written for INT28 (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 04-17; tests/fixtures/int28_compositions) and lane R1-CONSOLIDATE's w8a-dca-open (tests/fixtures/open_fill_order, same chart, same corpus bars). An instrumented copy of the adapter (scratch only) names the rules acting in each run: - int28-c1-two-parent-reversal-v6 (40 trades): W3's r1 cells every day at v6's defaults -- W3 rule C 9 times (a reversal voids the long's global exit) and W5 MR 24 times (a declined reversal leaves it the held long's); - int28-c2-pooc-global-pair-opening (8 trades): W3 rule B and W5 C2 at each of the eight commissioned process_orders_on_close openings, at the same fill; - int28-c3-exit-queue-explicit-pair-add (2 trades), a control: the KI-62 cover stands behind a queued explicit exit pair; - w8a-dca-open (18 trades), c4: R-B admits E3's P3 and E4b's HM and R-A fills E1's limits B, C, A, E2a's close_all before L1 and E4b's HM before HL; in E4b both act on one fill point. R-A's own test reads the R-B cells off TradingView's rows only (its tree has no R-B) and W8A's compares them in any order, so c4 pins the composed order, with every entry id. Fail-before, each built against the tree's own library and headers: on main 962960b and at the TVDEF-DROPS boundary c1, c2 and c4 fail ("25 passed, 4 failed"); at W3's last pick and at W8A's c4 alone fails ("27 passed, 2 failed": B, A, C in E1, L1 closed at once in E2a, HL before HM in E4b); from R-A's pick on "29 passed, 0 failed". W5's rules act in c1 and c2 without moving a trade TradingView books. R1-CONSOLIDATE's R1 (entry()'s same-bar batch, flush_pending_same_bar_commands() and resolve_terms()) acts beside W4 DORM, W4 F08 and W5 M2 on its own tapes (tests/fixtures/same_bar_reissue, all green on this tree); it only reverses a held position, while W6 F10's and W6B's pair rules, which share the batch's scope, act from flat. R-A leaves a flat book holding MARKET entries of both sides to the pair route, so on r1c-ra-prior-limit W6B's checkpoint acts in the pair cells and R-A in the others. W3B-ENG-GRID, picked right after W3, meets W4 F19a (resolve_terms() and fill_pooc_close_exits(): the fill-time size of a partial from_entry "" exit) and W5 MR (follows_same_bar_declined_reversal() and suspend_brackets_for_reversal(): a from_entry "" exit under a declined reversal). Both are composed in those picks (their INT28 notes say how) and pinned by W3B's own tapes w3bf05-g8 and w3bf05-r1c in tests/test_exit_queue_tapes.cpp, which fail on the uncomposed merge ("1022 passed, 2 failed") and pass composed ("1021 passed, 0 failed", W3B's own count); the header and the fixture README carry the reasoning. Not a composition, and not fixed here: a first version of c3 (AX2 a full exit created after an explicit qty=1 AX1 on the entry's bar) is booked by main and fails from W3 rule A's pick on -- rule A ranks exits by the kernel's acceptance order, and a full exit of a pending entry is accepted before an explicit-quantity one created before it. INT28's report carries the tape. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
W4-ENG-POOC-SAMEPASS rewrote the frozen legacy twins test_pooc_global_full_exit_l4c (F19a) and test_calc_on_order_fills_l4c (F19c) to TradingView's rows and recorded both in the twin-parity inventory, leaving the uncompiled base copies as they were. W5-ENG-MARGIN-V6, W6-ENG-FILL-ORDER and W8A-SIGSTATE-1 applied their observable rewrites to both copies; INT28 does the same for W4's, byte for byte (the twins' hunks apply unchanged to the base files), so every rewritten row reads the same in the pair. The checker reads the pinned base manifest, not these files: check_twin_parity.py: OK, unchanged. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… the integrated tree and attributed Nine pinned tables move on the integrated tree (main 962960b and the 50 wave-J picks). Each lane's own re-pin (TVDEF-DROPS 62d53f04, W3 16f1e370, W3B-ENG-GRID 7488d33e, W5 7c2753a5, W6's digests in d46096ec, W8D 7ed4ca0c, W8E's in bf4a0ba2, W4 9eac1f56, W8A d238a9d6, R1-CONSOLIDATE cd634e62 and 6b9a7ae8) was harvested on its own tree and is not applied; each table is re-harvested here once, with its TU's own switch, and the same harvest against main 962960b reproduces every old row of all nine (and of test_adapter_brackets_relower's R4 section, which does not move). The command allocations, calendar and host-view-memo witnesses do not move. W3B-ENG-GRID's void rule took the only quiet-bar run that reached the ThrottledReopens gate live, so test_adapter_quiet_bar_probe ("every gate taken both ways") needs the ThrottledStopEntry fixture that lane added to the battery in its re-pin; it is carried here, four new runs of the pinned table. Attribution: a harvest at every pick boundary (the 46 through R1-CONSOLIDATE R-A; 5ce0dd28 is ADR text only, and RUN-HARNESS's three picks change scripts, tests, docs and one pineforge.h declaration: a harvest at the last of them equals R-A's, table for table) marks each moved row with the pick(s) at which it changed, and an instrumented copy of the adapter (scratch only: a RULE marker wherever a picked rule decides differently from the code it replaced, folds new state, or, for the two compositions INT28 made in W5 MR's and W4 F19a's picks, takes the composed branch) shows the rule of each such pick firing in that run on the final tree. Where it does not, the pick is marked superseded by a named later pick, and the census at the boundary before that pick shows the rule firing in the run and the census after it does not: W3 rule A in 20 live-state digests, TVDEF-DROPS R3 in two (bracket seeds 1675665 and 2199310) and W3B's ledger booking in two quiet-bar runs (Random36 magnified, recording off and on) are superseded by W3B's void rule; W8D in chains seed 314189 by W8A R-B; W3B's ledger booking in Storm12 (magnifier off and on) by R1-CONSOLIDATE R1. - test_adapter_quiet_bar (PINEFORGE_QUIET_BAR_HARVEST): 216 of 272 runs move and 4 are new -- W3B void exit 188, W3B F02 ledger booking 138, W3B F02 first id 135, W4 F08 52, W8A R4 44, R1-CONSOLIDATE R1 26, W5 C1 21, W8A R-B 20, W3B g1 12, W4 F19a 10, W4 F19c 10, TVDEF-DROPS R2 8, TVDEF-DROPS R1 4, W3 rule C 4, W5 M1 4, W5 C2 4, W8E C1 4, W4 F14 4, W8A R2 4, W5 M2 2; 174 move at two picks or more. Its own margin-call floor goes 24 -> 10: the battery books 62 margin-call rows on main and after W3B, 54 after W5, 46 after W4's F08 to BRK and 10 from W4 F14 on (the entry-bar calls on the slipped percent shorts TradingView never opens). - test_publication_witness (PINEFORGE_D2D_PUBLICATION_HARVEST): 142 of 156 runs -- W3B void exit 112, W3B ledger booking 105, W3B first id 95, W4 F08 30, W3B g1 18, W5 C1 12, W8A R-B 8, R1-CONSOLIDATE R1 8, TVDEF-DROPS R1 6, W4 F19c 4, W5 MK 2, W8E C1 2, W4 MCSIZE 2, W8A R3 2, R1-CONSOLIDATE R-A 2, W8D 1; 105 move at two picks or more. - test_adapter_live_state_equivalence (PINEFORGE_V19E_HARVEST): 108 of 108 digests -- W3B void exit 108, W3 rule A 30, W8A R-B 26, W4 F08 21, W8A R3 11, W8E C1 5, R1-CONSOLIDATE R-A 5, W8D 4, W4 F19c 4, R1-CONSOLIDATE R1 4, TVDEF-DROPS R3 3, W5 C1 3, W4 F19a 1, W4 DORM 1; 71 move at two picks or more. - test_adapter_lookup_index_witness (PINEFORGE_P7_HARVEST): kFlip, kBrackets and kMargin -- W3B ledger booking 3, void exit 2, first id 1, W8E C1 1. - test_adapter_recording_hash_witness (PINEFORGE_F9_HARVEST): Submit (W3B first id and ledger booking) and Oca (W8A R-B); no trade moves. - test_adapter_receipts_in_place (PINEFORGE_P4_HARVEST): Revive, Groups and StopLimit (the last two magnifier off and on) -- W3B ledger booking 5, first id 2, W8A R-B 2, R1-CONSOLIDATE R-A 2 (three of Groups' trades). - test_adapter_report_relower (PINEFORGE_R1_HARVEST): 16 hashes of the three scenarios, W3B first id; no trade moves. - test_adapter_continuation_view (PINEFORGE_P1_HARVEST): 196 reads over nine tables, W3B void exit. - test_adapter_brackets_relower (PINEFORGE_R4D_HARVEST): rel-exit-before-entry's book digest, and by hand its leg counts {4,4,0} -> {0,0,0} (the test prints them), W3B void exit; its rows and equity do not move. Each table's note lists every moved row, value or declaration with its picks. Unmoved runs where a marker fires are the rules' no-effect cases (a margin check with no shortfall, a close call whose ledger booking matches the old reservation, a close pass with nothing to fill). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Counted with ctest -N per ci_verify profile, each build dir configured with ci_verify's own argv (configure only, no compiler launcher), on main 962960b and on the integrated tree: profile main INT28 floor release 724 748 RELEASE_MIN_TESTS 748 (no release row skips) kernel 294 296 KERNEL_MIN_TESTS 295 (the WebSocket row skips on a system libcurl) debug 705 729 EXCLUDED_REGISTERED_MIN 729 sanitizers 705 729 EXCLUDED_REGISTERED_MIN 729 native 714 738 EXCLUDED_REGISTERED_MIN 738 The same twenty-four rows join every profile but kernel, and none leaves: TVDEF-DROPS 1, W3-ENG-EXIT-ALLOC 1 (W3B-ENG-GRID adds to that row), W3B-ENG-GRID 1, W5-ENG-MARGIN-V6 1, W6-ENG-FILL-ORDER 1 (W6B-ENG-PAIRS adds to that row), W8C-SECURITY 1, W8D-NOTRADES 1, W8E-EXITS 1, W4-ENG-POOC-SAMEPASS 6, W8A-SIGSTATE-1 4, R1-CONSOLIDATE 2, RUN-HARNESS 3 and INT28's composition row; two of RUN-HARNESS's (test_run_strategy_magnifier, test_run_strategy_mincontract) are source-free and join kernel too. The lanes' own floor hunks, counted on their own trees, are not applied. The self-test's pinned table and docs/ci.md follow (scripts/test_ci_verify.py: 145 tests OK). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every anchor resolved on main 962960b (1725 of 1725). The picks changed no cited page but W8A-SIGSTATE-1 R-B's rows (design row MG3, the ADR-0001 max_open_lots row, the pine-to-native and contributor rows; merged at their pick with the integrated tree's digits), R1-CONSOLIDATE's two ADR-0001 keep_handle rows (no line anchors), RUN-HARNESS's ABI counts and runner notes (README.md, docs/coverage.md, docs/pages/abi-stability.md, docs/pages/index.md, the runtime cheatsheet; no line anchor moves with them) and docs/ci.md's floor paragraph (same line count). Each anchor into a file the picks changed is carried through an exact difflib line map from its origin's copy of the cited file to the integrated tree's (int24_exact_reanchor.py --origin 962960b --origin 5f1afd58 --origin 89fcbcd7 --exclusive-first; the second origin, R-B's pick before W3B-ENG-GRID joined the stack, decides only R-B's rows, and RUN-HARNESS's re-anchored tip decides none): 303 moved, 1410 kept, 12 stuck, set by hand from the same line map: - the KERNEL_MIN_TESTS and RELEASE_MIN_TESTS citations (CONTRIBUTING.md:155 and :156, contributing-llm.md:120, :121 and :267), whose lines the floor commit rewrote: ci_verify.py:293 -> ci_verify.py:301 and ci_verify.py:463 -> ci_verify.py:500; - max_open_lots (ADR-0001:708 and design MG3): 2199-2201 -> 2267-2269; - design MG3 accepted_in_cycle 6220-6269 -> 6387-6440 (R-B rewrote a line inside) and fifo_named_entry 8981-8987 -> 9390-9417 (W5 QP's basis lines); - design TR5 trail_one_shot 9798-9927 -> 10258-10347: the span ran 38 lines into the process_orders_on_close exit block W4 F08 deleted, and now ends where the retained code ends; - design SZ11 enqueue_pooc_fifo_close 7242-7592 -> 7588-7827: the span began in the close-reservation helper W3B-ENG-GRID retired; it now starts at enqueue_pooc_fifo_close and ends at the old end's image in flush_pending_closes. Digit changes only. check_doc_anchors: 1725 1725 0, "every documentation anchor resolves to its symbol". Carry check against the origins: 1689 cite the same text, 36 cite regions a pick rewrote (W3B-ENG-GRID's close ledger and void exit among them); stale carries 0. check_doc_lint: clean. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…gs to that reversal At Pine v6's defaults a default-size reversal entry is declined at the next open when its units at that open cost more than the equity they were sized from (declined_reversal_at_open). A strategy.close_all() placed on the same bar after that reversal entry belongs to it, as a strategy.close() of the held id already does (opposite_reversal_pair / the paired close): TradingView flattens the held position through neither. Where the open funds the reversal it closes the position and opens the other side, and the close_all does not flatten that side; where the open declines it the position is held and the close_all does not fill either. The lowering placed the close_all as a Flatten request of its own, which flattened the held position at a declined open. src/source/pine_adapter.cpp, close(), the empty-id (close_all) branch: the close_all is kept as a source observation (shadowed, as the empty-id entry pair's close_all already is) when an opening default-size entry of the other side was placed on the same script bar, under a percent_of_equity default of 100 or more, outside process_orders_on_close and calc_on_order_fills, on a fractional (or absent) lot grid. Pinned by five lab tv --no-note tapes of synthetic probes on BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-04-17 (tests/fixtures/ reversal_close_all), twelve cells each, eight of them declined at 100 %: int28fix-ca2-rev (reversal entry, close_all) and int28fix-ca2-frosty (job-2614's order: long entry, short entry, close of the held id, close_all) give TradingView's one tape, the same as the control int28fix-ca2-close-id (the pair the adapter already holds); the controls int28fix-ca2-first (close_all first: it flattens at the open) and int28fix-ca2-p99 (99 %: nothing is declined) bound the rule. tests/test_reversal_close_all_tapes.cpp replays them: 25 passed, 4 failed on 4efcb8c (both rule rows), 29 passed, 0 failed with the rule. tests/test_declined_reversal_close_leg_l4c.cpp R5 froze the close_all firing beside a declined all-in reversal (a characterization row, with no tape); it now holds the long, as these tapes do. The twin inventory records the observable rewrite, and INVENTORY_SHA256 is re-pinned. Hard-lane target job-2614-andrewwieiw-frosty-alerts on BINANCE:ETHUSDT.P 15 (TV-DEFAULTS codegen b33c9dc, lane replica of the lab verifier): its first divergence from TradingView on INT28 is the 2025-04-04 00:00 New York bar, whose end-of-evening strategy.close_all() follows the script's reversal entries; the 04:15 UTC open declines the reversal and the engine flattened the short. With the rule the run books TradingView's 40207 trades: weak 84.4 % / countAbsDelta 651 / pnl p90 144.8457 -> excellent 100 % / 0 / 0 (baseline weak 80.7 % / 3867 / 123.2267). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…nd their book is called once At Pine v6's defaults a script that places several same-side entries on one bar while flat -- each within the equity alone, together over it -- lost every entry after the first: validate_precommit judged each later fill on the book the earlier ones had just opened and refused it, and nothing was margin-called. TradingView, on synthetic probes (lab tv --no-note, BINANCE:ETHUSDT.P 15; tests/fixtures/margin_residual): - w5b-sb-market-60x2 (two explicit 60 % entries per cell, pyramiding 2): both fill at the next open and the 120 % book is called there, 4 x the lot-floored shortfall, FIFO (all of A, part of B); - w5b-sb-pooc-60x2: under process_orders_on_close both fill at the close and the book is called at the next open, sized at that close; - w5b-sb-default-x3 / -pooc-default-x3 (three default 100 % entries, pyramiding 3): all three fill and are called whole; where the next open gaps over the signal close all three are refused, as the first always was; - w5b-sb-market-40x2 (the control): an 80 % book fills with no call. The CG-LINUX-RED env_facts export of 2026-09-26 is the same shape. validate_precommit now admits a MARKET entry placed while flat that meets, at the point it fills, a book only its own bar's earlier flat market entries opened (flat_sibling_book) on its own cost against the equity it was sized from, the pyramiding cap still applied. A flat bar's priced entries keep their route: two flat-armed stops that fire on one bar pyramid past the cap (test_integration's probe-80 row). The opening checkpoint in on_applied waits for the last such sibling (flat_sibling_fill_follows) and then calls the combined book, a zero-fee all-in book included; under process_orders_on_close the call is taken at the next open (called_at_next_open), a long sized at the close as M2 sizes a short. tests/test_margin_residual_tapes.cpp (new) replays the five tapes: on the base 1ed0a9e4 the four rule rows fail and the control passes. test_pyramiding_count_differential P6 (three market entries on one flat bar at pyramiding 2) now books TradingView's two fills instead of three; its pinned adapter outcome moves to the tape's. Pinned hashes moved by this rule alone, re-harvested at this commit (the base reproduces every old row): publication witness 6 of 156, quiet bar 16 of 272, live-state equivalence 7 of 108. Doc anchors re-anchored on this tree. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit d6d7b7a9f6c7e87c7bae610b8e4f6e92c831a3b8) INT28-FIX: src/source/pine_adapter.cpp, on_applied's opening-margin checkpoint: this rule's `!sibling_book` and W6-ENG-FILL-ORDER F10's `same_point_pair_fill_follows(...)` (picked into INT28) guard the same condition; both stand. The pyramiding differential (M13 / G2-15) takes both lanes' moves: W8A-SIGSTATE-1's R-B moved P4, P4b and D1 to TradingView, this rule moves P6, so the adapter's count is 13 of 15 -- the test header, the fixture README, the ADR-0001 `max_open_lots` row and the design row MG3 say so. tests/CMakeLists.txt: git's clean merge put this lane's fixture if() inside INT28-FIX CA's block; each block is closed on its own. The digest pins (live_state, quiet_bar, publication witness) keep INT28's rows here and are re-harvested once on the integrated tree; the doc anchors likewise. INT28-FIX pick note: W5B-ENG-MARGIN-RESIDUAL 1 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…s checked over its path before the script A short carried under process_orders_on_close took no path margin check on a bar with a competing pending entry, nor on a rounded-money short (lot value under one unit, pyramiding <= 1): ab9714b's schedule sliced it after the script instead, at the bar's high, and deferred that slice behind the close's market fills to the book they left. So the script read the uncalled size, a close or a bracket that emptied the book first left no call at all, and an add filled at the close was called at that bar's high on the grown book. TradingView, on synthetic probes (lab tv --no-note, BINANCE:ETHUSDT.P 15, a fee- and slippage-free 100 % default short at each of fifteen signal closes; tests/fixtures/margin_residual): - w5b-pc-seen-size: at the next close the script buys back math.abs(strategy.position_size): the call at the bar's high comes first and the script reads the called size, so the buy closes exactly the rest; - w5b-pc-seen-size-parked: the same with a far long limit entry parked beside each short -- the same tv_trades.csv, byte for byte; - w5b-pc-carried-close / -carried-reverse: the call at the high, then the close's close_all, or the reversal sized from the called book's equity; - w5b-pc-bracket-tp: a take-profit the bar reaches after its high fills the rest of what the call at the high left; - w5b-pa-pooc-p50 (50 %, pyramiding 2): a short add filled at the close is not called at that bar's high; the grown book is called at the next open; - w5b-pc-carried-none (the control): nothing at the close, unchanged. schedule_margin_call_path now takes the path check on a carried POOC short beside a competing entry, on_bar_open no longer defers it for a rounded-money short (defer_rounded_pooc_short_margin_until_close is deleted, with its quiet gate), and on_bar_close drops the fee-free carried-short checkpoint and its deferral to the last close fill (on_applied). Commissioned or slipped shorts keep lane W5-ENG-MARGIN-V6's check before the script (C1). pooc_close_checkpoint_deferred_ms_ is never set now; it stays, as the constant it folds into the broker-state hash, so no digest moves for its removal. tests/test_margin_residual_tapes.cpp gains the seven rows: on the rule SB commit the six rule rows fail and the control passes. Behaviour pins that recorded ab9714b's schedule move to the tapes (legacy twin copies kept in sync; observableRewrites + INVENTORY_SHA256 re-pinned): - test_carried_pooc_short_margin_state: the competing-entry control reads -12.33168 (was -12.44432), the rounded-money one -12.44432 with two rows closed (was -12.60172, one); - test_rounded_carried_short_trail: the other order shapes and the competing entry see the margin before the script (-888216.89 / -884473.25, two closed), as the trailing exits already did; - test_l9a_carried_pooc_competing_checkpoint: every parked variant books its unparked control's rows. Pinned hashes moved by this rule alone, re-harvested at this commit (the SB commit reproduces every old row): publication witness 21 of 156, quiet bar 10 of 272, live-state equivalence 1 of 108. The MG15 anchor of native-feature-parity.md points at the exception left; doc anchors re-anchored. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 819b4ce75d409ca85c6f2f94d5e2995caed059e3) INT28-FIX: src/source/pine_adapter.cpp, schedule_margin_call_path(): this rule retires the legacy competing-entry predicate; INT28's picks had kept it, with W3B-ENG-GRID's void-call test in the resting_order scan that follows (a void call's request rests for no one). The retirement stands and the void_issue test stays. scripts/check_twin_parity.py: the twin inventory takes both INT28's and this rule's observable rewrites (key order), and INVENTORY_SHA256 is the combined file's. docs/design/native-feature-parity.md row MG15 takes this rule's chronology-exception sentence over INT28's anchor-only edit. The digest pins keep INT28's rows here and are re-harvested once on the integrated tree; the doc anchors likewise. INT28-FIX pick note: W5B-ENG-MARGIN-RESIDUAL 2 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ith no lot of slack An add to a held position of its own side kept one lot of slack: lane W5-ENG-MARGIN-V6's placement judgement (add_margin_exceeds_equity) and the fill-time add arm of validate_precommit both admitted an add whose held-plus-own margin exceeded the equity by less than one lot's notional, and the adapter then margin-called a whole unit of the book (the "family R" band). Under process_orders_on_close, where M1 does not judge at placement, the add was admitted at its close fill the same way. TradingView, on synthetic probes (lab tv --no-note, NYSE:F 15, whole shares; tests/fixtures/margin_residual): - w5b-pa-f-pooc-p50 (process_orders_on_close, 50 %, pyramiding 2): a long E1 at 09:30 .. 14:30 New York and an add E2 at the next close; every add whose margin on top of the held long's exceeds the equity at the close's tick is dropped -- by $1.90 .. $9.95 against a $9.9 .. $10.8 share as by more; - w5b-pa-f-market-p50: the market twin, judged at the signal close; - controls: w5b-pa-pooc-p40 (40 % + 40 % fits), w5b-pa-market-p50 and w5b-pa-pooc-p50-tick (BINANCE:ETHUSDT.P, 0.0001 lots). The margin-basis-frac probes of the population are the same shape. add_margin_exceeds_equity drops its lot term, and validate_precommit's frozen arm keeps one lot of slack for a flat opening only (same_side_add); the reversal arm had none already. tests/test_margin_residual_tapes.cpp gains the five rows and the NYSE:F bars (f15_bars.inc, the lab lane f-15 chart feed 80f404ae..., 2025-04-15 .. 2025-05-19): on the rule PC commit the two rule rows fail and the three controls pass. test_reversal_admission_float_guard pin E moves: its sub-lot shortfall add ($88 against a $106 lot) is now DECLINED (the legacy twin copy kept in sync; observableRewrites + INVENTORY_SHA256 re-pinned). No pinned hash moves (re-harvested at this commit). Every doc citation of the adapter is carried through the exact line map from the lane's base. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 32fae161f95f9d2f9e830d33e80453e2eff058b1) INT28-FIX: every code, test and fixture hunk applies unchanged. The twin inventory takes this rule's observable rewrite beside INT28's and INVENTORY_SHA256 is the combined file's. The docs keep INT28-FIX's composition of the pyramiding rows (W8A-SIGSTATE-1 R-B and this lane's SB); the doc anchors are re-anchored once on the integrated tree. INT28-FIX pick note: W5B-ENG-MARGIN-RESIDUAL 3 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… money, not the equity A default percent-of-equity quantity on an instrument with a lot grid was the lot floor of pct/100 * round10(equity): the adapter put the equity on TradingView's ten-significant-digit money grid before taking the percentage. TradingView rounds the order's money -- the percentage of the equity -- and floors that. At 100 % the two are one number; below it they part whenever the equity carries more decimals than ten digits keep. TradingView, on synthetic probes (lab tv --no-note, OANDA:EURUSD 15, 0.01 lots, initial_capital 10000000; tests/fixtures/margin_residual): - w5b-sz-eur-p50-10m and -p33-10m: every 4 bars from flat a default long filled at the close under process_orders_on_close, flattened two bars later; the equity rounded first floors 44 of the 504 quantities at 50 % and 37 at 33 % one lot off TradingView's, no rounding at all 8 and 18, the money rounded none; - the control w5b-sz-eur-p100-10m-0410: at 100 % both roundings reproduce all 130 quantities (no rounding misses 98). The population probe margin-basis-frac on OANDA:EURUSD (50 %) is the same shape: replayed from TradingView's first bar, all 3352 closed quantities match under the money rounded; under the equity rounded 1327 are a lot off. Rule PA exposed it there: an E1 one lot above TradingView's made the add TradingView admits exceed the equity by $0.007, and PA drops it (control data-OANDA-EURUSD/margin-basis-frac excellent -> strong on the PA head). default_sizing_cash (the non-cash-fee percentage) now rounds the money. The cash-fee branch (R5 lane PAR-CASHFEE's tapes) is unchanged. tests/test_margin_residual_tapes.cpp gains the OANDA:EURUSD chart (the lab lane eurusd-15 feed f945f31a..., 2025-04-01 .. 2025-04-29, eur15_bars.inc) and the three rows: on the commit before this one the two rule rows fail and the control passes; after, every row passes. No pinned hash moves (re-harvested at this commit). Every doc citation of the adapter is carried through the exact line map from the commit before. native-feature-parity's SZ3 row said "TV parts stay A: rounded equity `source_money_round`" (introduced by 2dd9834, R5 lane F2); it now names the percentage's money it rounds. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 0c11fd62ef860a54093ced6502a291e3e376d02f) INT28-FIX: every code, test and fixture hunk applies unchanged. The design row that names `source_money_round` takes this rule's sentence (the percentage's money is rounded, not the equity) over INT28's anchor-only edit; the docs keep INT28-FIX's composition of the pyramiding rows, and the doc anchors are re-anchored once on the integrated tree. INT28-FIX pick note: W5B-ENG-MARGIN-RESIDUAL 4 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…by one contract On a whole-unit lot grid TradingView follows a margin call of four times the floored restore with a call of one more unit at its next check point when the book's oldest lot is one unit and the first call left part of the book. The adapter took the first call only. TradingView, on synthetic probes (lab tv --no-note; tests/fixtures/margin_residual): - w5b-sb-es-fixed-n (CME_MINI:ES1! 15, point value 50, 1000000): every 2 hours 3 + cell % 5 one-contract MARKET longs from flat fill at the next open; a book whose restore floors to 1 or 2 contracts is called 4 or 8 at the open and one more at the bar's next path point (5641.5 open, 4 called there, 1 at the 5642.5 high); a sub-lot restore (one contract) or a call of the whole book takes none; - w5b-sb-es-pooc-fixed-n: under process_orders_on_close both at the next open; - w5b-es-short-q111111 and -pooc-short-q111111: shorts alike; - w5b-es-q1122, -pooc-q1122 and -q115 (1, 1, 5: the call ends inside the third lot): the oldest lot one contract, 4 then 1; - w5b-f-q1x12 (NYSE:F, 100 of capital, twelve one-share entries): 4 then 1 in all 70 cells; - controls: w5b-es-q21111, -pooc-q21111 and -q222 (oldest lot two contracts: 4, no follow-up) and w5b-eth-q111 (BINANCE:ETHUSDT.P, three 1 ETH entries on 0.0001 lots: 1.1848 called, none more). The population's CME_MINI:ES1! pf-probe-close-callsite-* probes (five one-contract process_orders_on_close longs, called 4 + 1 at the next open) are the same shape. whole_unit_follow_up_due decides it on the book before the call (the grid, a call of at least four units short of the book, the oldest open lot of one unit). The three places a flat sibling book (rule SB) is called take the follow-up: the market opening checkpoint in on_applied books it at the bar's next path point (next_source_path_waypoint, the forced execution price a path print books), on_bar_open books it at the same open for a process_orders_on_close long and short. No durable state is added. tests/test_margin_residual_tapes.cpp gains the CME_MINI:ES1! chart (the lab lane es1-15 feed 766e8149..., 2025-04-01 .. 2025-04-29, es15_bars.inc) and the twelve rows: on the commit before this one the eight rule rows fail and the four controls pass; after, every row passes. No pinned hash moves (re-harvested at this commit). Every doc citation of the adapter is carried through the exact line map; MG7 names the follow-up. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 38eee6c4a1d18b8670d29d2142adbcb2beba9ccd) INT28-FIX: every code, test and fixture hunk applies unchanged. The design row MG7 takes this rule's sentence (the whole-unit follow-up call) beside INT28's anchor-only edit; the docs keep INT28-FIX's composition of the pyramiding rows, and the doc anchors are re-anchored once on the integrated tree. INT28-FIX pick note: W5B-ENG-MARGIN-RESIDUAL 5 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…amiding A full-margin long of one lot takes TradingView's one-unit money call (the ten-significant-digit residual, submit_tv_money_long_margin_call). Under process_orders_on_close the adapter refused it whenever the strategy declared pyramiding above 1, a scope ab9714b pinned only for pyramiding 1. TradingView, on synthetic probes (lab tv --no-note, OANDA:EURUSD 15, initial_capital 10000000, 0.01 lots; tests/fixtures/margin_residual): - w5b-ms-eur-p100-pyr2: every 4 bars from flat one default 100 % long filled at the close, flattened two bars later, pyramiding=2: the same file as the pyramiding-1 w5b-sz-eur-p100-10m-0410, fourteen one-unit money calls included; the adapter booked none of them (122 trades against 130); - the control w5b-ms-eur-p99-pyr2 (99 %): no margin call at all, the adapter's 168 trades already TradingView's. The population probe margin-basis-allin on OANDA:EURUSD (100 %, pyramiding 2, process_orders_on_close) is the same shape: its first divergence is such a one-unit call the adapter does not book. The pyramiding term leaves the call's scope; the book must still hold one lot. tests/test_tv_money_carried_pooc_l4b.cpp's preserved-scope case pyramiding(2) -- a compatibility control that made no TradingView claim -- now takes the call carried() pins (the legacy copy kept in sync; observableRewrites + INVENTORY_SHA256 re-pinned). tests/test_margin_residual_tapes.cpp gains the two rows: on the commit before this one the rule row fails and the control passes; after, every row passes. No pinned hash moves (re-harvested at this commit). Every doc citation of the adapter is carried through the exact line map. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 3408c0ca5e6b378dabf67f00ea551431496cccf8) INT28-FIX: every code, test and fixture hunk applies unchanged. The twin inventory takes this rule's observable rewrite beside INT28's and INVENTORY_SHA256 is the combined file's; the docs keep INT28-FIX's composition of the pyramiding rows, and the doc anchors are re-anchored once on the integrated tree. This lane's floors commit (3c90030d) is not picked: the ctest floors are counted once on the integrated tree. INT28-FIX pick note: W5B-ENG-MARGIN-RESIDUAL 6 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…e its units fit at the signal close At Pine v6's defaults (100 % of equity, margin 100) a default-quantity percent_of_equity MARKET short is sized at the signal close less the slippage ticks, and the commission, which puts more units in it than the equity buys at the close itself. Under process_orders_on_close the adapter already refused such a short when those units' margin at the close exceeds the equity (lane W4-ENG-POOC-SAMEPASS, tests/fixtures/pooc_short_slip_admission). TradingView does the same when the short fills at the next open, whatever that open: it opens only while the units at the signal close, on its tick, fit the equity there. The adapter placed the short and filled it. TradingView, on synthetic probes (lab tv --no-note, NYSE:F 15, 2025-04-01 .. 2025-07-01; tests/fixtures/slipped_short): - int28fix-adm-s-s (10000, slippage 1): 6 of 124 cells open; all 10000 at a close of 9.905 sizes 1010 at 9.90 and does not open (1010 x 9.91 = 10009.1 over 10000); - int28fix-ou-s4 (4000, commission 0.04 %, slippage 4): none of 310 opens; - controls int28fix-adm-s-0 and -adm-s-c (no slippage: sized at the close, only the fill-time admission) and int28fix-adm-l-cs (a long, sized above the close) are unchanged. The lane's other probes (int28fix-adm-s-cs, int28fix-ou-s1, -s2, -s2b, -s3b and -s3, slippage 1 to 3; lane scratch) agree on every cell: 2479 cells, TradingView's decision and share count each, read at its own equity. src/source/pine_adapter.cpp, entry(): outside process_orders_on_close and calc_on_order_fills, a default percent_of_equity MARKET short with slippage placed from a flat book with no entry pending is dropped at placement when its frozen units at the tick-rounded signal close, times the short margin, exceed the sizing equity. A reversal and a book with an entry pending are not measured and keep their paths. tests/test_slipped_short_tapes.cpp replays the tapes (the SS rows compare TradingView's openings; their margin-call slices are rules OU and OF): 6138 passed, 2 failed on the commit before (both SS rows), 3530 passed, 0 failed with the rule. Leaver target job-2558-finnp17-silver-surfer on NYSE:F 15 (commission 0.04 %, slippage 1; TV-DEFAULTS codegen b33c9dc, lane replica of the lab verifier): its first divergence on INT28 is the 2025-04-03 14:15 short TradingView does not open; with this rule alone moderate 94.4 % / countAbsDelta 2 / exit p90 0.0058 / pnl p90 8.3262 -> moderate 94.1 % / 1 / 0.0041 / 4.4177, the rest of the gap being the entry-bar margin slicing rules OU and OF close. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…nit restores the book at its fill On a whole-unit lot grid a margin call whose restore floors below one lot takes one unit (the family R whole-drop band in source_margin_units). With slippage TradingView takes that unit only where it restores the book at the call's own fill: it prices the whole position's requirement at the liquidation's slipped print -- a short's buy is the slippage ticks above the mark, a long's sell below it -- against the equity at the mark, and closes nothing when that requirement exceeds the equity by a unit's margin or more. The adapter took the unit whenever the restore floored below one lot, on every all-in short with slippage: a one-share call at nearly every entry bar. TradingView, on synthetic probes (lab tv --no-note, NYSE:F 15, 2025-04-01 .. 2025-07-01; tests/fixtures/slipped_short): - int28fix-adm-s-s (10000, slippage 1): no one-share call at all; a short of 968 shares filled at 10.32, 9.12 short of its margin at the 10.33 print, takes none (968 x 10.34 exceeds the equity by 18.80); - int28fix-ou-s3 (4000, commission 0.04 %, slippage 3): a short of 382 filled at 10.43, 8.77 short at 10.46, takes none either; - controls int28fix-adm-s-c (no slippage: 22 one-share calls) and int28fix-adm-l-cs (a long's sell: 5) keep every call, as the formula gives them. Across the lane's other probes (int28fix-adm-s-cs, int28fix-ou-s1, -s2, -s2b and -s3b, slippage 1 to 3; lane scratch) and the leaver's own tape, a model of the entry bar with this formula, rule OF's follow-ups and, at slippage 2 and more, a short's check at the bar's low reproduces every call TradingView books there: 1671 entry bars. src/source/pine_adapter.cpp, source_margin_units(): the band's one unit now also requires the position's requirement at the liquidation's slipped print to exceed the equity at the mark by less than one unit's margin. At zero slippage the test is the band's own, and a long's sell always passes it, so slippage-free and long books are unchanged. The design row MG7 names the band's new condition. tests/test_slipped_short_tapes.cpp: the int28fix-adm-s-s row now compares TradingView's trades, not only its openings, and int28fix-ou-s3 joins: 3640 passed, 2 failed on the commit before (both rows), 3630 passed, 0 failed with the rule. Leaver target job-2558-finnp17-silver-surfer on NYSE:F 15 (lane replica of the lab verifier, codegen b33c9dc): with SS and this rule the engine matches TradingView's entry-bar calls up to its first one-unit follow-up (2025-05-21), moderate 94.1 % / countAbsDelta 1 -> 92.7 % / 4 (exit p90 0.0041 -> 0.0044, pnl p90 4.4177 -> 2.7382): the follow-ups TradingView books after these calls are rule OF, the next commit. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…is followed by one more unit On a whole-unit lot grid with slippage TradingView follows a one-unit margin call with one more unit at the bar's next path point when that unit did not restore the book at its own fill: the book it called, re-marked at the call's fill price, is still short of margin by more than a unit's margin and two slippage steps. The next point books the unit where its own check calls nothing and its print is inside the call's fill (below it, for a short's buy); a print at or beyond the fill is judged by its own check alone. A unit so taken is followed alike, and the close is the last point a bar offers. A long's call is never followed (its sell lowers the requirement with the equity), nor is a call at zero slippage. The adapter took the first unit only (W5B's follow-up, whole_unit_follow_up_due, covers a four-times call on a book whose oldest lot is one unit). TradingView, on synthetic probes (lab tv --no-note, NYSE:F 15, 2025-04-01 .. 2025-07-01; tests/fixtures/slipped_short): - int28fix-ou-s1 (4000, commission 0.04 %, slippage 1), through 2025-04-30: 29 follow-ups after 55 one-unit calls; a short of 352 shares filled at 9.86, called at 9.88 off the 9.87 open print and 0.068 short re-marked there, gives up one more at the 9.855 low; one of 359 filled at 9.63 and called at 9.65, 0.001 inside its margin re-marked there, takes none; one of 396 filled at 9.97 and called at 9.99 meets the 9.995 high at that fill, whose own check calls the next share, followed at the 9.94 low; - int28fix-adm-s-cs (10000, commission 0.04 %, slippage 1), through 2025-06-25: 3 follow-ups. The lane's slippage 2 and 3 probes (int28fix-ou-s2, -s2b, -s3b; lane scratch) place the threshold at two slippage steps: 0.038 short at slippage 2 takes none, 0.0445 takes one; 0.0439 at slippage 3 takes none, 0.21 takes one. Their replay also needs two shapes this rule does not model, left open in the fixture README with a third the leaver's tape holds. src/source/pine_adapter.cpp: follow_one_unit_margin_call() decides the follow-up after a one-unit call on the book the call left (the unit's own exit fee added back is the book it called, re-marked at its fill) and books it at the next path point through submit_margin_call_units, as W5B's follow-up does; on_applied calls it after the slipped market short's opening checkpoint and after every kernel liquidation of one unit. The design row MG7 names the follow-up beside W5B's. tests/test_slipped_short_tapes.cpp: rows int28fix-adm-s-cs and int28fix-ou-s1 join, each over the window that ends before the first unmodelled shape its tape holds: 6834 passed, 2 failed on the commit before (both rows), 6860 passed, 0 failed with the rule. Leaver target job-2558-finnp17-silver-surfer on NYSE:F 15 (lane replica of the lab verifier, codegen b33c9dc): with SS, OU and this rule the engine books TradingView's 481 trades, strong 97.9 % / countAbsDelta 0 / entry p90 0 / exit p90 0 / pnl p90 0 (INT28 moderate 94.4 % / 2 / 0 / 0.0058 / 8.3262; baseline strong 96.5 % / 10 / 0 / 0.0264 / 99.8631). Its remaining divergences are calls on the later bars of a held position (the close-to- open shape among them). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Counted with ctest -N per ci_verify profile, each build dir configured with ci_verify's own argv (configure only, no compiler launcher), on INT28 4efcb8c and on this tree: profile INT28 INT28-FIX floor release 748 751 RELEASE_MIN_TESTS 751 (no release row skips) kernel 296 296 KERNEL_MIN_TESTS 295 (unchanged) debug 729 732 EXCLUDED_REGISTERED_MIN 732 sanitizers 729 732 EXCLUDED_REGISTERED_MIN 732 native 738 741 EXCLUDED_REGISTERED_MIN 741 The same three rows join every profile but kernel, and none leaves: test_reversal_close_all_tapes (rule CA), test_margin_residual_tapes (W5B-ENG-MARGIN-RESIDUAL, picked) and test_slipped_short_tapes (rules SS, OU and OF). W5B's own floors commit is not applied. The self-test's pinned table and docs/ci.md follow (scripts/test_ci_verify.py: 145 tests OK). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…e integrated tree and attributed Three pinned tables move on the integrated tree (INT28 4efcb8c and INT28-FIX's ten commits). W5B-ENG-MARGIN-RESIDUAL's own re-pin was harvested on its own tree and is not applied; each table is re-harvested here once, with its TU's own switch, and the same harvest against INT28 4efcb8c reproduces every old row of all thirteen tables. The report re-lowering, recording hash, continuation view, receipts, command allocations, lookup index, host-view memo, calendar and both brackets re-lowering witnesses do not move. Attribution: a harvest at every commit boundary (CA, SB, PC, PA, SZ, FU, MS, SS, OU, OF) marks each moved row with the commit(s) at which it changed. Only the picked W5B rules SB and PC move any row; CA, PA, SZ, FU, MS, SS, OU and OF move none, and no row moves and comes back. - test_adapter_quiet_bar (PINEFORGE_QUIET_BAR_HARVEST): 26 of 276 runs -- W5B SB 18, W5B PC 10 (Random16 recording on, magnifier off and on, at both). - test_publication_witness (PINEFORGE_D2D_PUBLICATION_HARVEST): 27 of 156 runs -- W5B PC 21, W5B SB 6. - test_adapter_live_state_equivalence (PINEFORGE_V19E_HARVEST): 4 of 108 digests -- W5B SB 3, W5B PC 1. Each table's ledger names the moved rows' old and new values. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The picks and this lane's rules move lines in src/source/pine_adapter.cpp, include/pineforge/source/pine_adapter.hpp and scripts/ci_verify.py that the pages, the ADR and the design notes cite. check_doc_anchors.py --fix placed the citations whose symbol occurs once; reanchor_doc_citations.py carried 112 more across the line map from INT28 4efcb8c (where every anchor resolved) and W5B-ENG-MARGIN-RESIDUAL's tip 3408c0ca (for the sentences its picks wrote). Seven were set by hand: RELEASE_MIN_TESTS ci_verify.py:507 (three citations), MG4's AdmitWithHostMargin span 13755-13856 (the INT28 map; the W5B tip's reading of a row that predates it was not taken), MG7's whole_unit_follow_up_due 15276, and MG15's `4.0 * minimum` 14505 and resting_order 15375-15390. Every re-anchored span starts and ends on the lines it did at 4efcb8c; the four that change length hold code inserted inside them (W5B's picks and rule OU). check_doc_anchors: every documentation anchor resolves to its symbol; doc lint and doc reverts are clean. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…'s finer request.security values TradingView runs a calc_on_order_fills recalculation of a historical bar on the whole bar: a recalculation reads the bar's close, high and low, and a request.security on a timeframe finer than the chart reads the bar's own slice of it, as the bar's close calculation does. The native scheduler fed a chart bar's slice of the auxiliary finer feed only in its close callback, after the bar's fill recalculations, so a recalculation read the chart bar's close, high and low but the previous chart bar's finer values. (ab9714b fed the slice before dispatch_bar, ahead of the bar's fills.) TradingView, on three synthetic read-out probes (lab tv --no-note, BINANCE:ETHUSDT.P 15, 2025-04-01 .. 2025-05-01; tests/fixtures/coof_recalc_security): a recalculation closes the position with a comment naming what it read. - int28fix-rc, the open fill's recalculation, 2025-04-11 00:30 UTC: close 1520.73, high 1523.56, low 1514.04 (the bar's own) and a lookahead_off 5m close of 1520.73 (the 5m bar closing with the chart bar); the engine read 1519.95, the previous bar's. The same on the 02:30 short (1538.66; the engine read 1545.13). - int28fix-rc2, the recalculation of a take-profit filled on the bar's path (1520.95): the same values. - int28fix-rc3, the open fill's recalculation: a lookahead_on 5m site reads the bar's first 5m bar (1521.55) and a 5m request.security_lower_tf array the bar's three (1521.55 .. 1520.73). src/source/pine_scheduler_native.cpp: PineScheduler::recalculate feeds the bar's auxiliary slice at the bar's first recalculation before its close callback (callback_advances_source_bar), through scheduler_feed_aux_security, which rebases the script-state checkpoint on the fed values; PineScheduler::bar feeds it only when no recalculation did. A recalculation after the close callback (a fill the close books, strategy.close(immediately = true)) finds the slice that callback fed and feeds nothing: fed again, the finer sites step twice, and on the run's last bar the slice's routing is already cleared. The design row HT4 names the rule. tests/test_coof_recalc_security_tapes.cpp: ports of the three probes over the corpus 15m and 1m bars, checkpointing their sites' values and the 5m site's completed-bar count as a generated strategy does, require every trade and every read-out the tapes carry to be the engine's, and every callback of every bar to read that bar's own 5m close with its slice fed once; a fourth probe closes positions immediately at a bar's close, mid-run and on the last bar. 158 passed, 38 failed on the commit before (every recalculation reads the previous bar's slice); 196 passed, 0 failed with the rule. Gated on the first recalculation alone (without the close-callback guard) the fourth probe fails 12 of 194: from the mid-run close on, the 5m site is three bars ahead, and the last bar's run fails ("auxiliary request.security chart routing is not initialized"). Leaver LV3, finnp17-cleantradequantum-v7-2-ob-fvg-integrated on NASDAQ:AAPL 15 (lane replica of the lab verifier, codegen 2440321): strong 99.0 % / countAbsDelta 2 / entry p90 0 / exit p90 0 / pnl p90 0 (INT28 moderate 94.9 % / 3 / 0 / 0.0038 / 0; baseline strong 95.1 % / 1 / 0 / 0.0037 / 0). The script closes a short in its open fill's recalculation when a finer request.security turns its signal long; reading the previous bar's values, the engine never did, and since W4 F19c books the through stop that recalculation places at the open, the stop took the trade instead of the close and the reversal the script set with the close was lost. Layer: source-adapter parity runtime (ADR-0001, Decision, layer 2); no kernel unit changes. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Counted with ctest -N per ci_verify profile, each build dir configured with ci_verify's own argv (configure only, no compiler launcher), on 3332347 and on this tree: profile 3332347 INT28-FIX floor release 751 752 RELEASE_MIN_TESTS 752 (no release row skips) kernel 296 296 KERNEL_MIN_TESTS 295 (unchanged) debug 732 733 EXCLUDED_REGISTERED_MIN 733 sanitizers 732 733 EXCLUDED_REGISTERED_MIN 733 native 741 742 EXCLUDED_REGISTERED_MIN 742 The one row that joins every profile but kernel, and none leaves: test_coof_recalc_security_tapes (rule RC). The self-test's pinned table and docs/ci.md follow (scripts/test_ci_verify.py: 145 tests OK). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
RC adds two lines to PineScheduler::bar and twelve to PineScheduler::recalculate (src/source/pine_scheduler_native.cpp), and the floors commit four comment lines above RELEASE_MIN_TESTS (scripts/ci_verify.py), so the anchors below them moved: - check_doc_anchors.py --fix: coof_recalculation_due 611 -> 613 and PineScheduler::applied 606 -> 608 (docs/design/native-feature-parity.md); - reanchor_doc_citations.py --ref 3332347 (line map, each result verified on this tree): kFirstOpenLoopGuard 729 -> 743 (ADR-0001 and two design rows) and PineScheduler::recalculate 636 -> 638 (two design rows); - by hand, where the line map found no reference placing the symbol: RELEASE_MIN_TESTS ci_verify.py:507 -> :511 (CONTRIBUTING.md, docs/pages/contributing-llm.md twice). check_doc_anchors.py: every documentation anchor resolves to its symbol; check_doc_lint.py, check_design_inventory.py, check_kernel_seam_rows.py, check_dangling_comment_names.py: 0 findings. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
INT28: wave-J engine integration (paired with codegen-oss TV-DEFAULTS)
This PR integrates the accepted wave-J engine lanes onto main 962960b: 50 picked commits, each carrying
-xand anINT28 pick note: <lane> <n>trailer, plus 5 integration commits. It merges together with codegen-oss PR #147 (TradingView's Pine v6 defaults for an omitted capital and order size). The parity gate measures the two as a coordinated pair.Lanes
strategy.exit; exits armed before an opening; global exit voiding on reversal.keep_handlerows).syminfo.mincontractcomes from the lane's quantity step.Integration commits
test_int28_rule_compositions, 29/0).Verification (head 4efcb8c)
ci_preflightpineforge/parity, from a sweep of this head paired with codegen-oss fix(engine): fill stale prior-bar exit brackets marketable at a reversal fill bar's open #147.🤖 Generated with Claude Code
https://claude.ai/code/session_01LifAW9e5J6gFx7gatuxETE