ГЛОБАЛЬНЫЙ РЕФАКТОРИНГ АРХИТЕКТУРЫ - Рефакторинг этапов Stage 3.3: убран шаблонный LIMITED_WITH_REASON и переведен в контекстный мягкий отказ с полезными next-step.

docs/TECH/1CLLMARCH-FACT.md

@@ -2382,6 +2382,25 @@ Implemented in current pass (Stage 3.2 semantic route arbitration):
    - Stage 3 focused suite: `10` files / `76` tests passed.
    - Type build: `npm --prefix llm_normalizer/backend run build` passed.
 
+Implemented in current pass (Stage 3.3 soft-refusal + anti-template limited replies):
+1. Reworked address-lane limited reply composer from fixed template to contextual soft-refusal:
+   - Added deterministic phrasing variants (stable, non-random) to reduce repeated boilerplate.
+   - Kept concise "Коротко" structure while replacing hard repeated wording.
+2. Added context-aware explanation and recovery offers for limited responses:
+   - Response now injects request scope hints when available (organization, as-of date, period window).
+   - Added "Что могу сделать сейчас" with nearest supported scenarios (documents/payments by anchor, open contracts/tails, account-balance drilldown).
+   - Added weak-anchor suppression for user-facing hints to avoid low-value auto-substitutions in suggestions.
+3. Improved unsupported aggregate handling UX:
+   - For aggregate/ranking style queries, reply now proposes evidence-first path (collect factual base -> continue in extended analysis) instead of hard static fallback phrase.
+4. Improved missing-anchor UX:
+   - Missing anchor tokens are normalized to user-facing terms (e.g., `counterparty_or_contract` -> "контрагент или договор").
+   - Added concrete reformulation hint for underspecified anchor cases.
+5. Regression updates:
+   - Updated `addressQueryRuntimeM23.test.ts` soft out-of-scope assertion to validate contextual soft-refusal structure.
+6. Validation snapshot:
+   - Extended regression pack: `11` files / `344` tests passed.
+   - Type build: `npm --prefix llm_normalizer/backend run build` passed.
+
 Acceptance (Stage 3):
 1. LLM outputs strictly validated schema for extraction/decomposition (no free-form).
 2. Deterministic guards can block or downgrade answers when evidence insufficient.