fix: per-archetype PremortemTask decomposition for small-model compatibility

PremortemAnalysis required the LLM to emit a deeply nested schema:
3 AssumptionItem + 3 FailureModeItem, with 11+ required fields per item
including linked cross-reference IDs. Qwen 3.5-35B and other small local
models consistently echoed the schema structure back instead of producing
values, causing validation errors after exhausting all retries.

Fix (applies the 'LLMs handle narrative; code handles structure' principle):
- New ArchetypeNarrative schema: 5 plain text fields only (no IDs, no
  cross-references, no counts). The LLM writes narrative content only.
- Per-archetype decomposition: one independent LLM call per archetype
  with up to 5 retries. Failed archetypes are skipped gracefully.
- IDs, indices, and cross-references (assumption_id, failure_mode_index,
  root_cause_assumption_id) are assigned by code, not the LLM.
- falsifier derived from test_now field to avoid hardcoded tautologies.
- _calculate_risk_level_verbose returns 'Not Scored' when likelihood or
  impact is None (was rendering 'Likelihood None/5, Impact None/5').
- Dead code removed: ArchetypeAnalysis class, PREMORTEM_SYSTEM_PROMPT_TEMPLATE.

Validated: PremortemTask PASSED on GLM 4.7 Flash (HVT_minimal run).

Author: EgonBot