fix: reset compiled script runs and align MFI source direction - #223

Reusing a compiled strategy handle could retain Pine series, indicator and collection state across backtests. Add an engine-owned preparation hook after broker reset and before precalculation/security evaluation, and reset run-local allocation counters and cached broker snapshots. The companion codegen change resets generated persistent members through that hook. Configured inputs survive a new run, while ticks and order-fill recalculations within one stream retain their state.

MFI also classified equal-decimal source values by raw binary comparison, allowing a tiny HLC3 rounding residue to contribute a full bar's money flow. Use the existing Pine float-comparison helpers for direction. Direct TradingView controls distinguish below-band and above-band movements at three price scales. The fixes have no strategy, symbol or date lookup.

Addresses engine issue #219. This pair requires regenerated strategy C++ and rebuilt modules against matching engine headers/library; the internal C++ namespace prevents stale objects from silently binding the changed virtual layout. The public C ABI is unchanged. Companion codegen commit: c8ffe587c8fb82435fec34121a93fab2b7018924.

Validation:

• Final Cloud sweep cand-r37-full-b-20260909: 72/72 cases, 4,190/4,190 probes; 4,180 excellent / 10 strong / 0 moderate, up one excellent from Round 36. The README-inclusive pair reproduces the first full sweep's grades, quantities and trade CSVs exactly.
• Ford 15m RSI/MFI becomes excellent: canonical match 95.2% to 100%, count gap 2 to 0. EURUSD MFI and Ford EWO remain excellent and reach 100% canonical match.
• All 704 hard probes retain identical grades, quantity metrics and trade CSVs; 4,187/4,190 CSVs are unchanged. No verifier, grading, threshold, profile, feed, reference, input or population changes.
• Independent raw audit: 315/315 strict entry-and-exit matches across the three changed probes, 13 gained and none lost. Overall verifier matches 2,817,481 of 2,819,967 TradingView trades.
• Real retained-C-handle Cloud diagnostics on fixed 5,000-bar windows: Cashflow 42/41/41 becomes 42/42/42; CHoCH/BOS 133/134/134 becomes 133/133/133. SMA152 stays at 232 trades with repeated-run hashes now equal. Fresh controls agree; these are current windows, not the reporter's older counts.
• Full native build: 249/249 checks. Full codegen suite with candidate engine headers/library: 2,802 passed, 1 pre-existing skip. Permanent compiled C-ABI tests cover repeated/short batches, input A-B-A, stream restart/continuity and mutable collections/UDTs with order-fill recalculation. New literal MFI controls fail on the old source (17 failures) and pass all 157 checks on the candidate.
• Independent Grok 4.6 high review of the exact final engine/codegen pair: zero actionable P0/P1/P2 findings. Runtime and tests are unchanged from the fully tested pair; only README documentation was added before the final Cloud sweep.

Known limits: the EURUSD open-position mark retains its pre-existing price difference, and CSV display-precision differences remain. Exact MFI behavior at the inclusive 1e-10 boundary is inherited from the existing TV-pinned generic helper; the new direct MFI controls bracket that boundary. No all-column or byte-for-byte reporting equality is claimed.

Cloud PR gate pineforge-pr-gate-xh6xb: PASS, target score +1, hard regressions 0, tolerated exceptions empty, full 4,190-probe coverage. Verdict recorded and corroborated in Postgres for engine 0d152028f540bf48242ea190cf33144756ba6b40 and codegen c8ffe587c8fb82435fec34121a93fab2b7018924. Candidate snapshot 8bd08fd50eb8b1785ad364dd7759e9a5e4b0625e9f3b2c13e64816d77788d52c (788,836 bytes), unchanged population 3d72f815de54e4a2f6fdff6bb2294ea99c38849888390ded4ecd68ac3e5a9110. Independent final audit verifies all 72 immutable case/input manifests, source bundles, 64/64 runner tasks and successful pipeline completion. Both companion PRs must merge before the composite baseline promotion.