Fix durable lean investigation repair feedback and context validation

docs/investigation-lean-v19.md

@@ -33,3 +33,7 @@ Das Paket speichert kurze Beobachtungen samt LLM-Referenz, bewertetem Ergebnis u
 Lokale Fixture-Tests prüfen Schema/Cache-Roundtrip, verzögerte fachliche Updates, Wechsel-/periodische Konsolidierung, Fortsetzung per defer, Iterations- und Budgetende, ungültige Tools, ungelesene Referenzen, portable Wiederaufnahme und Unterbrechung nach Evidenzpublikation. Bestehende Investigation-/Gateway-/Early-Diagnosis-Tests sichern den vorherigen Modus ab.
 
 Die Offline-Messdatei replayt alte v18-Kontexte mit leerer v19-Sitzung. Sie misst Zeichen, keine Tokens oder tatsächliche neue Untersuchung. Laufkosten und Denkarbeit müssen einschließlich der Konsolidierungen in einem neuen kostenpflichtigen Lauf gemessen werden; während der Implementierung wurde kein LLM-Dienst aufgerufen.
+
+## Korrekturrückmeldungen
+
+Schemafehler werden vor dem Speichern auf den begrenzten Korrekturvertrag abgebildet. Ungültige Befund-IDs erhalten den konkreten Hinweis auf `B0001` (B plus mindestens vier Ziffern). Bereits gespeicherte Kontexte mit dem früheren `validation_keyword` bleiben lesbar; für den nächsten Prompt wird daraus ein sicherer Korrekturhinweis, ohne das alte Paket zu verändern. Neue Promptkontexte werden vor ihrer Veröffentlichung validiert.

src/viano_bug_analysis/analysis/investigation_compact.py

@@ -239,7 +239,8 @@ def validate_tool_shape(context,decision):
         if error:
             from viano_bug_analysis.llm.diagnostics import schema_feedback
             from .investigation_config import InvestigationError
-            raise InvestigationError('INVESTIGATION_OUTPUT',feedback=schema_feedback(error,include_keyword=True))
+            from .investigation_lean import repair_feedback
+            raise InvestigationError('INVESTIGATION_OUTPUT',feedback=repair_feedback(schema_feedback(error,include_keyword=True)))
         from .investigation_lean import checkpoint,build
         if checkpoint(context):
             import json

src/viano_bug_analysis/analysis/investigation_lean.py

@@ -93,6 +93,8 @@ def build(context):
         document['input']['tool_schema']['request_contract']=schema(context,strict=True)['properties']['next']
         full['messages'][0]['content']=files('viano_bug_analysis.llm.prompts').joinpath('investigation_lean_v19.txt').read_text(encoding='utf-8')
     full.update(prompt_version=VERSION,output_schema=contract)
+    if 'previous_rejection_details' in document['input']:
+        document['input']['previous_rejection_details']=repair_feedback(document['input']['previous_rejection_details'])
     full['messages'][1]['content']=model_json(document)
     return full
 
@@ -121,3 +123,21 @@ def normalize(context, reply):
     result['update'].pop('next_step',None)
     result['update']['confidence']=context['hypotheses'][context['selected_hypothesis']]['confidence']
     return result
+
+
+def repair_feedback(details):
+    """Project schema failures to the durable, value-free repair contract."""
+    result=copy.deepcopy(details)
+    keyword=result.pop('validation_keyword',None)
+    path=result.get('validation_path','')
+    import re
+    if keyword=='pattern' and re.fullmatch(r'/update/working_brief/findings/[0-9]+/id',path):
+        result['requirement']='use_stable_finding_ids_B_followed_by_at_least_four_digits_example_B0001_not_case_fact_ids'
+    elif keyword=='required' and 'path_prefix' in result.get('validation_fields',[]):
+        result['requirement']='provide_missing_required_fields; path_prefix="" searches_repository_root'
+    elif keyword is not None:
+        result.setdefault('requirement',{
+            'required':'provide_missing_required_fields','additionalProperties':'remove_unrecognized_fields',
+            'type':'match_expected_type_at_validation_path','enum':'use_allowed_value_at_validation_path',
+            'const':'use_allowed_value_at_validation_path'}.get(keyword,'match_full_controller_contract_at_validation_path'))
+    return result

src/viano_bug_analysis/analysis/investigator.py

@@ -165,7 +165,9 @@ class Investigator:
                 raise InvestigationError("INVESTIGATION_GENERATION_CHANGED")
             state["parent_ref"]=reader.manifest["artifacts"].get("investigation_state") if previous else None
             state["sequence"]+=1
-            if context is not None:state["active"]["context_sha256"]=digest(context)
+            if context is not None:
+                validate(context,"investigation-context")
+                state["active"]["context_sha256"]=digest(context)
             if state["sequence"]>self.limits.max_state_checkpoints:raise InvestigationError("INVESTIGATION_HISTORY_LIMIT")
             ref=f"analysis/investigation/{state['analysis_revision']:04d}/states/{state['sequence']:08d}.json"
             validate(state,"investigation-state")

src/viano_bug_analysis/schemas/investigation-context-v1.schema.json

@@ -464,6 +464,11 @@
           "type": "integer",
           "minimum": 0,
           "maximum": 1000000000000
+        },
+        "validation_keyword": {
+          "type": "string",
+          "maxLength": 64,
+          "pattern": "^[A-Za-z][A-Za-z0-9]*$"
         }
       }
     },

tests/test_investigation_lean.py

@@ -183,3 +183,58 @@ def test_replay_after_tool_publication_does_not_repeat_model_or_budget(tool_case
     assert result.status=='PROGRESSED' and len(adapter.calls)==1
     assert len(state(result)['hypotheses']['H1']['lean_session']['turns'])==1
     assert len(BudgetManager(tool_case[0]).snapshot(*CASE).state['accounts']['H1']['actions'])==1
+
+
+
+def test_bad_consolidation_finding_id_repairs_without_corrupting_context(tool_case,tmp_path):
+    from viano_bug_analysis.case.contracts import validate
+    class BadIdOnce(LeanAdapter):
+        failed=False
+        async def generate(self,config,request,*,max_response_bytes):
+            response=await super().generate(config,request,max_response_bytes=max_response_bytes)
+            c=self.calls[-1]['input']
+            if self.calls[-1]['task']=='consolidate_investigation_session' and not self.failed:
+                self.failed=True
+                eid=next(iter(c['evidence']))
+                response.document['update']['working_brief']['findings']=[{'id':'F01','direction':'neutral',
+                    'text':'Schema available.','scope':'Captured schema.','evidence_refs':[eid]}]
+            return response
+    adapter=BadIdOnce();router=InvestigationToolRouter(project_research=service_for(tool_case))
+    loop=Investigator(tool_case[0],gateway(adapter),InvestigationConfig(lean_loop=True),router=router)
+    first=run(loop,steps=1)
+    before=BudgetManager(tool_case[0]).snapshot(*CASE).state
+    rejected=asyncio.run(loop.step(*CASE))
+    assert rejected.status=='REJECTED'
+    assert BudgetManager(tool_case[0]).snapshot(*CASE).state==before
+    with PackageReader(rejected.package) as reader:
+        c=json.loads(reader.read_bytes(state(rejected)['active']['context_ref']))
+        validate(c,'investigation-context')
+        detail=c['previous_rejection_details']
+        assert 'validation_keyword' not in detail
+        assert 'B0001' in detail['requirement']
+        # The original released version persisted this extra keyword. Its
+        # immutable context must remain loadable without a package rewrite.
+        c['previous_rejection_details']={'validation_path':'/update/working_brief/findings/0/id',
+            'validation_fields':[],'validation_keyword':'pattern'}
+        validate(c,'investigation-context')
+        prompt=build_prompt(c)
+        assert 'B0001' in json.loads(prompt['messages'][1]['content'])['input']['previous_rejection_details']['requirement']
+    portable=CheckpointStore(Workspace(tmp_path/'portable'));portable.intake(rejected.package)
+    final=run(Investigator(portable,gateway(adapter),router=router))
+    assert final.status=='COMPLETED' and PackageValidator().validate(final.package).valid
+
+
+
+def test_invalid_context_is_not_published(tool_case):
+    from viano_bug_analysis.case.errors import CaseBuildError
+    adapter=LeanAdapter();loop=Investigator(tool_case[0],gateway(adapter),InvestigationConfig(lean_loop=True),
+        router=InvestigationToolRouter(project_research=service_for(tool_case)))
+    result=run(loop,steps=1);saved=state(result)
+    # A new prompt context is validated before a checkpoint can reference it.
+    saved['active']={'hypothesis_id':'H1','iteration':2,'attempt':0,'phase':'prompt',
+        'context_ref':'analysis/investigation/0001/contexts/invalid.json','context_sha256':'0'*64,
+        'decision':None,'llm_ref':None,'test_request':None,'test_deadline':None,'retry_token':'initial'}
+    before=tool_case[0].latest(*CASE)
+    with pytest.raises(CaseBuildError) as caught:loop._save(saved,context={'unexpected':'invalid'})
+    assert caught.value.code=='INPUT_SCHEMA'
+    assert tool_case[0].latest(*CASE)==before