Harden structured LLM outputs and add preflight CLI input schema vali…

[RichardDawald]

Addresses two reliability gaps:

1. LLM structured-output robustness at critical nodes (especially traceability), where half-structured JSON could break the chain.
2. CLI input quality guardrails, by validating --input payloads locally before backend requests.

The goal is fail-soft on model output and fail-fast on bad CLI payloads.

1. Traceability node: secondary structural repair + minimal field completion

Changes:

• Updated traceability_check_node to add a second-stage repair path:
  • Run normal agent.run_structured(...).
  • Post-process output with a deterministic repair function to ensure minimum usable structure.
  • If structured generation/validation still fails, return a deterministic minimal ClaimTraceabilityReport instead
    of failing the whole draft chain.

Added helpers:

• _repair_traceability_min_fields(report_payload, claims_payload)
  • Ensures each claim has a traceability report row.
  • Ensures each report has at least one elements_evidence.
  • Fills missing/empty fields with safe placeholders.
  • Preserves schema compliance via ClaimTraceabilityReport.model_validate(...).
• _build_minimal_traceability_report(claims_payload, failure_message)
  • Deterministic fallback report when output is unrecoverable.
  • Marks items as Unsupported and requests human review in risk assessment text.
• _extract_claim_rows_for_traceability(claims_payload)
  • Best-effort extraction of claim numbers/elements for fallback generation.
  • Includes synthetic fallback row if claims are missing.

Behavioral intent:

• Significantly reduce end-to-end failures caused by half-structured model JSON at traceability stage.
• Keep workflow progressing with explicit “needs human review” semantics rather than hard-fail.

2. CLI --input preflight schema validation (before request)

Changes:

• Added local validators:
  • ensure_nonempty_string_field
  • ensure_array_of_nonempty_strings_field
  • ensure_object_field
• Added validate_cli_input_schema(workflow, payload) for workflow-specific checks:
  • draft: require idempotency_key and (disclosure_text or disclosure_file_id)
  • oa: require idempotency_key and (oa_text or oa_notice_file_id), optional shape checks for related fields
  • compare: validate optional key types and require meaningful input signal (file ids/paths)
  • polish: require at least one of application_file_id / original_claims / application_specification
• Hooked validation into execution path:
  • after payload assembly (build_cli_request_payload)
  • before any API call (post_json)
• Error message prefix:
  • CLI input schema validation failed: ...

Behavioral intent:

• Early, readable local failures for malformed --input.
• Fewer invalid backend requests and clearer operator feedback.

Why This Is Safe:

• Changes are additive and defensive.
• No API route signature changes.
• No normal-path degradation: valid model outputs and valid CLI payloads keep existing behavior.
• Fallback paths are only activated under malformed model output or invalid CLI inputs.