fix(engine): stop retro-arming explicit-zero-offset trails from the carried peak - #148

Two trailing-exit rules diagnosed from TradingView tapes (TV is ground
truth). Family: winthetrade-ema-9-vwap-strategy-with-atr-trailing-stop
(strategy.exit("x", trail_points = atr*2, trail_offset = atr*2) -- a
price-unit ATR passed as TICKS) on 7 lanes: matchPct 100, entry/exit
timestamps exact, exit PRICE wrong on the same bar, so PnL / equity /
qty diverge from there.

(B) The bar's open is the first price a resting trail observes.

resolve_exit_path_fill only folded prices into the trail state AFTER
each O->L / O->H segment (update_exit_trail_state). With a POOC close
fill the carried best is the entry price itself (apply_market_order_fill
same_bar_close_fill guard), so a trail whose activation level the next
bar's open already sits past stayed dormant through an adverse-first
leg, armed only at the favourable extreme and filled at extreme -/+
offset. TV's emulator arms at the open with best = open and fills the
retrace at open -/+ offset.

  OANDA:XAUUSD 15m long 2025-04-04 10:45Z @3110.31; 11:00Z bar
    O 3110.40 H 3136.775 L 3109.24 C 3134.46, atr*2 ~ 15 ticks
    (activation 3110.325, offset 0.015): TV "Long Exit" 3110.385 =
    open - 0.015 (PnL 24.47); engine 3136.76 = high - 0.015 (PnL 8631).
    194 of the 195 differing XAUUSD exits fit TV = open -/+ k ticks vs
    engine = extreme -/+ the same k, always adverse-leg-first
    (|O-L| < |H-O| for longs) with the open beyond the activation.
  NASDAQ:AAPL 15m short 2025-04-02 18:45Z @222.93; 04-03 13:30Z bar
    O 205.54 L 202.52: TV 205.55 (open + 1 tick), engine 202.53.
  NYSE:F 1D long 2025-04-21 @9.47; 04-22 O 9.55 H 9.72: TV 9.54,
    engine 9.71.

Fix: after try_exit_open_gap_fill declines (same liveness condition --
path cursor at the open, non-entry bar unless magnifier / cascade
waypoint), observe_exit_trail_open folds bar.open into the running best
and arms the trail iff the OPEN itself is past the activation. It never
arms from the carried best, so the #148 one-shot rule (an explicit zero
offset never retro-arms) is untouched; for exit-at-activation shapes an
open past the level was already gap-filled, so it is a pure best-price
observation there.

(A) A trail_offset in [0, 1) ticks is TV's explicit-zero one-shot trail.

compute_exit_trail_state gated the zero rule on trail_offset != 0.0 and
only then floored, so an offset in (0, 1) became a FINITE zero distance:
exits_at_activation = false and the stop rode exactly on the running
extreme, filling at the bar's best price. TV floors first and applies
the zero-offset rule: fill at the activation crossing, or at the open
when the bar opens past it, and no carried arming.

  Pinned with `lab tv` on OANDA:EURUSD 15m (2025-04-01 -> 05-01): the
    tapes for strategy.exit("x", "L", trail_points=3, trail_offset=0),
    trail_offset=0.5 and trail_offset=0.9 are byte-identical
    (190 rows, sha256 36aa80ac...).
  EURUSD 15m short 2025-03-31 03:45Z @1.08330, atr*2 ~ 0.0006 "ticks"
    (activation ceil -> 1 tick = 1.08329, floor(offset) = 0); exit bar
    O 1.08330 L 1.08314: TV 1.08329 (the activation), engine 1.08314.

Fix: floor(trail_offset) first; a zero-tick result takes the existing
explicit-zero path (trail_offset_price NaN, one_shot_zero_offset_trail).
Whole-tick and larger fractional offsets keep the nils123456-orb /
legalrice floor rule unchanged; negative offsets are untouched. The
trail-peak fold in engine_fills.cpp already floors, so it agrees.

Tests: tests/test_trail_open_arm_subtick_offset.cpp (105 checks) pins
the XAUUSD / AAPL / F shapes at the resolver and, for XAUUSD and AAPL,
end to end through a POOC close-fill entry; 0 == 0.5 == 0.9 == 0.6 at
the activation, at a gapped open, and for the #148 no-retro-arm hold
(resolver and engine level, with the omitted-offset durable-arming
control); whole-tick 1.0 / 1.4 keep the floored trailing distance.
Against the pre-fix source 29 of those checks fail, all of them the
tape cells.

One existing pin contradicted the TV rule and is re-pinned, not
dropped: test_path_resolve_extra.cpp fl_sub (0f1f06e, a factorial
round, 2026-08-25) asserted that a 0.6-tick offset "rides the extreme
itself once armed (distinct from an explicit 0)" -- fill at the peak
102.00 -- as an extrapolation of the floor rule with no tape behind
the sub-tick cell. It now expects the activation 101.00 and equality
with the explicit-0 fill, with the `lab tv` evidence in the comment.
Every other trail pin (test_path_resolve_extra, the #148 retro-arm
cells, test_margin_call_trail_exit_chronology D, cascade G4) passes
unchanged. ctest 153/153 (was 152/152 before the new binary), scripts
unittest 102/102.

Co-Authored-By: Claude Fable 5.1 <[email protected]>
Claude-Session: https://claude.ai/code/session_014vizwdJMsDkHsjmS4StvZY

…rs; re-pin the one-shot fixtures

test_zero_offset_trail_rides: 32 `lab tv` tapes (NYSE:F 15m feed
80f404ae85ef, NASDAQ:AAPL 15m feed ae2b03d3736f) replayed end to end
through BacktestEngine, each under the sensor's own issue mode (re-issued
per bar, once, alongside the entry, or the probe's close*1.5%/mintick
re-issue), asserting TV's exit price and bar.

Re-pinned to the tape-confirmed rule: test_trail_fill_snap (the explicit-0
AAPL 196.135 open is a path level fill at position 0, still 196.13
end to end), test_path_resolve_extra and test_trail_open_arm_subtick_offset
(the #148 fixtures now carry the best family Z actually carries — the
issuing close — and HOLD; a carried best past the activation arms every
zero-tick offset, the contract the command layer's restart relies on; a
gapped open arms the trail and it rides), test_margin_call_trail_exit_
chronology fixture D (the explicit zero arms from the placement close like
the omitted offset). ctest 190/190.

(cherry picked from commit 5f87ccb)