fix: preserve reversal carry and pinned COOF exit behavior

[luisleo526]

Preserve the one-unit closing carry for the pinned default-100% reversal after a same-signal close-point margin call. Let a carried full COOF exit reach an outward-rounded chart extreme while preserving the raw path cursor. Settle already-resting long stop siblings on one adverse segment before one script recalculation, retaining each fill price and broker fill event.

Validation on engine dd98f02ed492f64ae8b48358117c83530f88ae74 with codegen 3fd97fe28abd7b191cb376d9f5db4e056fcfae4b:

• Cloud Run completed 72/72 cases and measured all 4,190 probes. EURUSD taro and ES1 daily jenmb move from strong to excellent; all other 4,188 probes retain their canonical metrics. Zero metric regressions, tier drops, engine errors, coverage gaps or new moderate results.
• Gate pineforge-pr-gate-68lz2: PASS, zero regressions or tolerated exceptions across 704 hard probes, target score +2. Result: 4,154 excellent / 35 strong / 1 moderate.
• Ford JOAT trade 30 now exits May 29 at 13:45 UTC instead of 14:00, matching TradingView. This corrects the trade timestamp without changing its grade.
• Fresh all-target build, 208/208 C++ unit tests, C ABI and freshness checks passed. Independent Grok 4.6 review of this exact combined commit is GREEN, P0/P1/P2 = 0.

The fixes retain their TradingView-pinned order and execution-mode guards. EURUSD taro retains PnL P90 of 0.0007; ES1 daily jenmb has zero canonical P90 errors. Grading metrics, thresholds, tolerances, verifier, harness and population are unchanged.

Candidate snapshot: 23fc951fbb02d9f38ca4875ef1aceccb5bf570b79285651afb1419ad399eb917 (788,848 bytes), compared with baseline engine 76518c6b79cea8410b462a892994deaebd4bdc9a.