fix(loop): use string_length instead of first_char to detect valid response

[hlgreenblatt]

Summary

Loosen the safety-net check at src/loop.metta:65 from (== "(" (first_char $resp)) to (> (string_length $resp) 1). The original first_char == "(" check fails reliably for valid Python-string returns from py-call even when the returned string obviously starts with (, because of a String-vs-atom type mismatch in the SWI-Prolog↔MeTTa binding for first_char/2.

When this check fails, the loop substitutes (REMEMBER:OUTPUT_NOTHING_ELSE_THAN: ((skill arg) ...)) for $response and the agent's actual output gets discarded — the agent sees only the reminder string in the next iteration. The bug surfaces any time a skill body calls a Python helper that returns a string, including any new skill that wraps py-call and returns an s-expression.

The length-based replacement is permissive enough to let well-formed Python-string returns through and still gates against the empty-string failure mode that the original check was guarding.

Provenance

This bug is documented in a downstream fork's implementation notes from 2026-04-25 (gotcha #3 in a dual-format-LLM-adapter work item). The same one-line fix is in production use in that fork. Submitting upstream as a standalone single-concern PR.

Test plan

• Existing agent loop continues to function normally — empty-string responses still trigger the reminder branch.
• Skills that return strings via py-call (any new Python-bridged skill) now have their output reach $response cleanly instead of being dropped.
• A simple regression: define (= (echo $s) (py-call (helper.normalize_string $s))), ask the agent to call (echo "(send hi)"), observe that the next iteration's RESULTS contains (send hi) instead of REMEMBER:OUTPUT_NOTHING_ELSE_THAN:.

Scope

One line, one file. No new dependencies. No behavioral change for valid (...)-prefixed responses (the new check is strictly more permissive on those). The only behavioral difference is that single-character outputs are now treated as valid and routed to sread, where they will fail parsing cleanly via the existing HandleError MULTI_COMMAND_FAILURE_... path rather than being substituted for the reminder string.