Classify UDF-wrapped and var-scalar stable expressions as TA lengths

[luisleo526]

What

The stable-TA-length guard hard-rejected two bar-invariant shapes:

1. A TA constructor length computed through a pure single-expression user-defined function — any UDF call was classified as series-valued, so e.g. ta.lowest(low[1], f_bars(mins)) with f_bars(m) => math.max(1, math.round(m*60/tfSec)) refused to transpile.
2. A never-reassigned var scalar with a stable per-run initializer (e.g. var tfSec = timeframe.in_seconds()) — all var names were treated as mutable state.

The stability classifier and the reset-expression renderer now inline single-ExprStmt-body UDFs structurally (parameter substitution over the arithmetic/call grammar; cycle guard + depth cap at the shared choke-point, so recursive UDFs are refused without hanging), and never-reassigned top-level var/varip scalars with stable initializers register as stable derived expressions (no use-site inlining). Genuinely series-dynamic lengths (ternary over an ATR-derived ratio) are still rejected with the same error — pinned by a guard test and an end-to-end true-positive sentinel.

Gates (author + independent fresh-context R7, both green)

• 3 RED tests fail pre-fix with Unsupported TA constructor length, pass post-fix; 4 conservative guards (series-dynamic, reassigned-var body, recursive, multi-statement) pass on both sides.
• Full pytest 1473 passed / 1 skipped; corpus compile sweep 254 passed.
• End-to-end: the false-positive exemplar transpiles and runs (89.5% match, 100% price parity); the true-positive sentinel still errors on its series-dynamic length.
• Private validation full sweep (412): transpile-error inventory 8→7 (exactly the unblocked exemplar), remaining errors byte-identical, 0 tier moves.

🤖 Generated with Claude Code