fix(codegen): resolve input-derived TA lengths and input.timeframe request.security tf

[luisleo526]

Summary

Two related transpiler resolution gaps where a structural argument derived from an input (not a literal or direct input alias) silently degraded the generated indicator:

1. Computed/parameter TA lengths → silent period-1. A ta.* length that is arithmetic-of-input (wilderLen = rsiLen*2-1, wp = sf*2-1) or a function parameter (f(src, len) => ta.sma(src, len)) fell back to period 1 (unsmoothed). Now folded to the compile-time value / expanded to the real runtime length expression; genuinely unresolvable lengths raise a guardrail instead of silently emitting 1. Function per-call-site TA clones now get globally-unique member names, so a helper reached via multiple enclosing functions no longer emits duplicate member declarations.
2. input.timeframe()-backed request.security timeframe → bare/empty. tf = input.timeframe("60") then request.security(sym, tf, …) emitted tf as a bare variable, uninitialized at evaluator-registration time, so the HTF series degraded to the chart timeframe. Now resolved to its runtime input value.

Validation (TradingView deep-backtest, BINANCE:ETHUSDT.P @ 15m)

• 3commas DCA cohort: tao 0→100%, dual-rsi 0→100%, bb-rsi 23→99%, sol 7→93%.
• QQE pair: qqe-trend 1.4→99.8%, qqe-channel 18→84%; jayentriken-bbwp 84→88%.
• Scraped set verify_corpus tiers: excellent+strong 42→57 / 90.

Corpus parity (hard gate) — intact

• verify_corpus.py --all = 251 excellent / 1 documented TV anomaly, 0 weak/moderate/minimal.
• Only 2 corpus strategies' generated C++ changes (mtf-htf-confluence-manual-trail, …-static-bracket): both moderate/weak → excellent, realigning codegen with the committed validated output (HEAD had drifted). The other 252 are byte-identical.
• Full pytest: 1372 passed, 1 skipped (+6 new regression tests). 0 new transpile/compile failures across the scraped set.

🤖 Generated with Claude Code