fix(skill_engine): parse LLM JSON despite trailing brace prose

[santhreal]

Problem

_extract_json used a greedy \{.*\} match, so valid JSON followed by prose that contained braces failed to parse and analysis/capture returned None.

Repro

from openspace.skill_engine.analyzer import ExecutionAnalyzer

text = '{"task_completed": true}\n\nReplace {foo} with {bar}.'
assert ExecutionAnalyzer._extract_json(text) == {"task_completed": True}

Fix

Decode from the first { with json.JSONDecoder().raw_decode, matching agents/base.py. Same change in capture_semantic._extract_json.