fix(rag): Unknown knowledge-database configuration keys are silently accepted and stored

[tandau-bbv]

The Form-duality contract introduced by #1822 promises that a knowledge database's configuration is validated against the form its ingestor announced. Validation of declared fields works, but fields the ingestor never declared pass straight through and are persisted, so a mistyped knob is accepted, stored and silently ignored at ingestion time. Found while end-to-end verifying #1841.

Current behaviour: POST /knowledge/databases/{database} accepts a configuration containing keys no announced form element declares, and writes them verbatim into BucketEntity.configuration. IngestorConfig sets extra="allow" (packages/core/swiss_ai_hub/core/ingestors/ingestor_config.py:28), so the model built by ModelCreationService.create_config_model and checked by InstanceConfigHelper.validate_config_for_create (packages/api/swiss_ai_hub/api/routes/knowledge/knowledge_service.py:455-457) retains unrecognised fields rather than rejecting them.

Required fields are still caught — a typo that omits llm_model correctly fails with llm_model: Field required. The gap is optional knobs: the three enrichment flags, vision_model, and any optional field a custom pipeline declares.

Expected behaviour: a configuration carrying a field the ingestor's announced form does not declare is rejected with a 400 naming that field, matching #1822's acceptance criterion.

Repro:

1. Create a database with a typo on an optional flag and an optional model:

POST /api/v1/{tenant}/knowledge/databases/kbtypo
{"ingestor":"document_ingestion","configuration":{
  "name":{"en":"Typo"},"description":{"en":"d"},
  "llm_model":"text-generation/Kimi-K2.6","embedding_model":"embedding/bge-m3",
  "with_table_refinemnt":false, "vision_modl":"text-generation/MinerU2.5-2509-1.2B"}}

2. Observed: HTTP 200, and the row stores both misspelled keys.
3. The pipeline then resolves the defaults for the fields the user meant to set:

user typed with_table_refinemnt=false  ->  actual with_table_refinement = True
user typed vision_modl=MinerU...       ->  actual vision model          = text-generation/Kimi-K2.6

The user believes they disabled an expensive enrichment step and chose a vision model; neither took effect, and the stored configuration looks deliberate. The same permissiveness lets one ingestor's knobs be submitted to another (crawl_depth from a custom ingestor is accepted by document_ingestion), weakening the per-ingestor isolation #1822 claims.

In scope

• Reject a knowledge-database configuration that carries fields the database's ingestor did not announce, with a 400 naming the offending field.
• Keep the existing behaviour for rows already stored: a database created before this fix, or whose pipeline has since dropped a field, must keep resolving and ingesting unchanged.

Out of scope

• The same extra="allow" on AgentConfig (packages/core/swiss_ai_hub/core/agents/agent_config.py:98), which likely lets agent instances accept undeclared config keys too. The Form stack is shared, so a fix here may generalise — but the agent/process surface needs its own verification and issue rather than being changed blind.
• Capability filtering of the vision-model picker — see Capability- and location-aware model selection for agents #1769 (sibling).
• The retriever honouring configuration["embedding_model"] — see feat(rag): Retriever reads the embedding model from the knowledge database #1820 (sibling).
• Any path to edit a stored configuration after creation; today there is none, so this issue only concerns the create path.

Accepted when

• Creating a knowledge database with a misspelled optional field is refused, and the message names that field.
• Creating a knowledge database with a field belonging to a different ingestor is refused.
• A valid configuration for the database's own ingestor is still accepted, including omitted optional fields.
• Databases created before the fix continue to ingest with their stored configuration, unchanged.
• make test passes in the affected scopes.