Fix Transformer_NonRecursive corrupting the tree on a non-last Discard

[chuenchen309]

Problem

Transformer_NonRecursive corrupts the tree when a callback returns Discard for a node that is not the last child processed. It slices its internal stack by len(x.children), but discarded children were never pushed onto that stack, so the slice over-consumes and pulls an unrelated sibling into the wrong parent:

from lark import Tree, Token
from lark.visitors import Transformer, Transformer_NonRecursive, Discard

class Drop:
    def drop(self, children): return Discard

tree = Tree('start', [Tree('keep', [Token('T', 'x')]), Tree('wrap', [Tree('drop', [])])])

class TBase(Drop, Transformer): pass
class TNR(Drop, Transformer_NonRecursive): pass

TBase().transform(tree)   # Tree('start', [Tree('keep', [Token('T','x')]), Tree('wrap', [])])   (correct)
TNR().transform(tree)     # Tree('start', [Tree('wrap', [Tree('keep', [Token('T','x')])])])       (WRONG)

keep is silently dropped from start and re-parented under wrap. Transformer_NonRecursive is documented as "Same as Transformer", and the base Transformer (unaffected) is the correct twin. A differential run of 3000 random trees using Discard gave 307 mismatches vs the base Transformer before this change, 0 after.

Cause

In Transformer_NonRecursive.transform, size = len(x.children) counts all original children, but a child whose callback returned Discard was never pushed (if res is not Discard: stack.append(res)), so stack[-size:] reaches into earlier siblings.

Fix

Push discarded results as placeholders so every node consumes exactly len(children) slots, and filter the placeholders when building a parent's args. A wholly-discarded tree now returns None instead of mis-unpacking an empty stack.

Existing tests pass (test_transformer_variants only exercised Discard on a last child, which masked this); added test_transformer_variants_discard_with_preceding_sibling covering all four transformer variants. mypy clean.

Note on #1614

As a side effect of the stack-alignment fix, a fully-discarded root now returns None — the same case #1614 targets for the crash. The interior-sibling corruption fixed here is distinct (it survives #1614's if not stack guard), but there is overlap on the root case; happy to rebase around #1614 or adjust if you'd prefer to keep them separate.

---

Disclosure: authored by an AI coding agent (Claude Code) running on this account — it found the bug, wrote the repro/fuzz, the fix and this description. I review every change and am accountable for it; the verification above is real and re-runnable from the diff. If this isn't a change you want, say so and I'll close it.