Classify UDF-wrapped and var-scalar stable expressions as TA lengths - #54

…s the member's key (C3 W4)

var n = input.int(9) with no title: the member reads the input under "n" (a
var initializer lends its name to the untitled call; the manifest lists "n"),
but the TA runtime reset read it under "", so overriding n never resized
ta.ema(close, n). Since 097f499 (#54, which turned the shape from a refusal
into an accepted TA length), _collect_stable_var_scalars recorded the var's
initializer as a derived expression, and the reset re-lowered that spelling
as an inline input call with no enclosing name. #133 spells that call with
its title and keyword arguments, which fixed the titled form, but an
untitled call still has no name to key by.

A never-reassigned var bound straight to an input call holds the input's
value on every bar, exactly like n = input.int(9), so it is now registered
the same way (_collect_known_var): input-backed, read through its call node.
Every reader then keys it by _get_input_title(call, var_name=n), and the
constructor placeholder is the default (9) as for the plain binding instead
of 1. The titled var binding takes the same path; var n = input(9), refused
before, is admitted with it. An input whose default is not a constant keeps
the derived-expression path.

No corpus probe binds an input with var: the 312 re-transpile byte-identical.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…the member and the manifest (C4 W4)

TradingView: "If not specified, the variable name is used as the input's
title" (Pine v6 reference, input.int title). The manifest and the member
followed that rule, but several getters re-derived the key from a
re-spelled call or from whichever name they were handed, so an override
set the member and never reached the TA size or the security feed:

- tf = input.timeframe("60") as a request.security timeframe registered the
  evaluator with get_input_string("", ...) -- since 6545234 (#29), which
  renders the bound call with no enclosing name;
- len = input.int(9) with a stable `len := ...` reassignment: the TA reset
  read the folded `(cond ? 30 : input.int(9))` under the default's text "9"
  from 8390364 on, under "" since C2's spelling;
- var n = input.int(9) * 2 (or math.max(input.int(9), 2), or a float
  feeding ta.bb): the reset re-read the nested call under "" -- 097f499 (#54);
- a = input.int(9) then b = a: the reset read it under "b" -- since the
  initial public release 5b92828.

One key per input, from the call node itself:
pine_spelling.input_binding_names() maps each global-scope input call a
declaration names -- the call bound straight to `v = input.*()` and any call
in a var initializer, the manifest's rule -- to that name.
_get_input_title() consults it before the name its caller passes, so the
member, the TA reset, a security timeframe and an alias's read all agree
with the manifest; and a call re-spelled for a derived length (codegen
_arith_expr_to_str, analyzer _expr_to_str) carries the name as title=, since
the re-parsed node has lost its declaration. The manifest walker moves to
pine_spelling.global_input_calls() so both share one definition of "global
scope".

E2E (tests/test_e2e_untitled_input_keys.py): each untitled shape runs next
to the same strategy with an explicit title, at the defaults and under the
same override keyed by each one's manifest title; the trades must be
identical and the override must move them. Before (on d0ae5be): 7 failed --
security_timeframe(_alias), stable_reassignment, var_nested(_call,_float),
alias -- each "override left the trades unchanged"; the direct binding,
input.source, a security length and a ta.change length through an alias
passed. After: 11 passed (e.g. security_timeframe override {"tf": "240"}
5508 trades == the "TF"-titled twin).

Re-transpile: no C++ or manifest change in the 312 probes, the corpus or
the 101 closed strategies -- none uses these shapes.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>