docs(confidence-engine): record focused deconstruction repeatability

docs/current-handoff.md

@@ -1904,3 +1904,100 @@ The current handoff captures all operational facts needed to resume today. For h
 **Next boundary:** rerun exactly one substantive compound-question answer-deconstruction observation through the corrected production route.
 
 **Zero live model calls occurred during the fix.**
+
+## Focused-deconstruction repeatability evidence — post-fix
+
+### Plumbing state (valid production path)
+
+```
+focused prompt
+→ focusedDeconstructJsonSchema supplied to Ollama /api/chat
+→ provider response wrapper
+→ .response unwrapped
+→ focused-deconstruction validator
+```
+
+The previous focused-deconstruction 502 observations before these fixes are **invalid semantic evidence** because they were contaminated by:
+
+1. provider-envelope misuse (semantic fields buried under wrapper);
+2. initial-reconstruction schema being supplied to the focused-deconstruction call.
+
+### First valid post-fix semantic observation
+
+One substantive compound-answer run returned:
+
+```
+HTTP 200
+schema valid
+supplier evidence preserved
+weekend-shift uncertainty preserved
+epistemic separation preserved
+unsupported inference: NO
+steering: NO
+follow-up relevance: PASS
+```
+
+### Repeatability observations (3 controlled Qwen/Ollama runs)
+
+Three additional/controlled identical Qwen focused-deconstruction runs were executed with the same fixed question and answer.
+
+| Metric | Result |
+|---|---|
+| Qwen/Ollama calls | 3 |
+| HTTP 200 | 3/3 |
+| schema valid | 3/3 |
+| supplier evidence preservation | 3/3 PASS |
+| weekend-shift uncertainty preservation | 3/3 PASS |
+| epistemic separation | 3/3 PASS |
+| unsupported inference | 0/3 |
+| steering | 0/3 |
+| follow-up relevance | 3/3 PASS |
+
+**Representation variance was minimal:** associative relationship present in some runs and absent in another; follow-up wording varied; core epistemic meaning remained stable. Do not claim deterministic wording or topology.
+
+### Product interpretation
+
+> For this fixed compound supplier/weekend-shift case, Qwen's initial compression did not prevent the deeper focused-investigation stage from repeatedly recovering the important epistemic distinction once substantive user evidence was supplied.
+
+This supports tolerating some initial representation compression when the underlying meaning remains available and later focused deconstruction can recover independently investigable epistemic states. **This is not yet generalised to a production invariant.**
+
+### Explicitly not claimed
+
+```
+initial decomposition quality no longer matters
+all compound questions are safe
+Qwen is generally reliable
+Qwen is production-ready
+Terra is unnecessary
+exact graph topology never matters
+the prompt is solved
+the product is commercially ready
+```
+
+The evidence is one fixed semantic pattern — a lower bound, not a warranty.
+
+### Current architectural lesson
+
+```
+deterministic code owns state / validation / graph integrity
+
+LLM owns semantic interpretation and deconstruction
+
+cheap direct question surfacing remains acceptable when meaning is obvious
+
+when meaning requires interpretation, focused LLM deconstruction is the semantic recovery layer
+```
+
+No new deterministic word lists, dictionaries, or semantic heuristics.
+
+### Live calls during this evidence task
+
+- Qwen/Ollama: 0
+- Terra/OpenAI: 0
+- Application API: 0
+
+### Git
+
+**Commit message:** `docs(confidence-engine): record focused deconstruction repeatability`
+
+---