Skillscript

Upgrade impact: none (additive). Two cases that previously errored now succeed; nothing that already worked changes behavior.

Both fixes come from Perry dogfooding a list-driven enter-project skill (thread 9ed7554b):

• foreach over an empty/whitespace-only string input now iterates zero times (was: one iteration with an empty element). Passing RULE_IDS="" into a skill and iterating it used to run a single pass with RID="", which then blew up downstream (e.g. amp_get_memory("")). An empty string is "no items", so it now matches [] and an absent ref — both already yielded zero iterations. A JSON-array string and a non-empty scalar are unaffected (still parsed / still wrapped to one iteration).
• A |fallback now rescues an unresolved reference regardless of its position in the filter chain. Previously only a lone |fallback, or a |fallback placed first, rescued an undefined base; any transforming filter ahead of it (${x|trim|fallback:"d"}, ${x|length|fallback:"0"}) threw Unresolved variable reference before the fallback could fire. Now an undefined base propagates lazily past intervening filters to the fallback — matching Jinja's undefined|filter|default contract, so a chain containing |fallback never explodes on a missing ref. Scoped to undefined only: a value that resolves — including a whitespace string like " " — still flows through every filter unchanged (${WS|length|fallback:"0"} with WS=" " is still "2", not "0"), so no existing behavior shifts. A chain with no |fallback still throws on an unresolved ref exactly as before.