Reused pf_strategy_t is not pure for compiled scripts: series/indicator state survives reset_run_state() #219

Found while building pineforge-live's ABI v4 binding (Plan B1 Task 1, 2026-09-08).

Symptom. Calling run_backtest_full twice on the same pf_strategy_t with the same bars yields different trade lists: run 2 has one extra/missing trade near the start. Reproduced on three corpus strategies at engine main 41c9c74 (ABI v4): ta-sma-152-close-cross-01, analyzer-cashflow-cash-costs-01 (34 vs 33 trades), analyzer-parity-choch-bos-isolator-01 (125 vs 126). Two FRESH handles (strategy_create each) agree byte-for-byte, and the per-bar broker-state hashes match on fresh handles.

Cause (hypothesis). reset_run_state() resets the broker/engine state but not the compiled script's own series/indicator state (e.g. var variables, KahanWindowSum buffers in generated.cpp), so the second run starts "warm" and an indicator becomes ready earlier.

Why the existing pin missed it. tests/test_live_flags_off_identity.cpp (PR #218) pins reused == fresh on a synthetic modulo-schedule strategy with no indicator state, so it is vacuous for this class.

Asks.

1. Either reset compiled-script state at the top of every run() (codegen + engine hook), or document pf_strategy_t as single-use per run in pineforge.h (strategy_create per run).
2. Add a reused-handle pin over a corpus strategy with indicator state (e.g. run ta-sma-152-close-cross-01 twice on one handle and compare trades + hashes with a fresh handle), so scripts/live_flags_off_identity.py's evidence and the live runtime's assumption are covered.

pineforge-live works around it at the binding level (fresh strategy per run, ordered setter replay) until this lands.