v2.26.1

Fix routine MCP request logging privacy, reject invalid encrypted DNS views,
and correct collection status, capsule determinism, and validation packaging.

Tool Surface Changes

Tool surface changes: no CLI command or flag changes. Maintainer validation
helpers add provider selection and endpoint overrides.

Security

• Suppress MCP server request-level logging by default (runtime level WARNING)
  to protect investigated domains from leaking into client log files. Full
  request logging is opt-in via RECON_MCP_LOG_REQUESTS=1. Preserve root
  logging configuration during MCP SDK construction.
• Reject typos and invalid values in RECON_DNS_VIEW on CLI lookups with exit
  code 2, preventing quiet fallback to unencrypted host DNS when an operator
  intended encrypted resolution.

Fixed

• Prevent offline lookups from exiting 0 when old certificate summaries without
  subdomains exist in cache: when all live DNS queries fail and no services are
  observed, a cached certificate summary without subdomains no longer recovers
  the DNS source, correctly exiting 4 (all_sources_failed). (#312)
• Disentangle certificate transparency status on domains with no certificates:
  soft-empty provider responses without recovered certificates or cache are no
  longer recorded as ct_provider_used, eliminating contradictory degraded and
  recovered status messages. (#313)
• Canonicalize lexical observation example order: sort matching subdomain
  labels alphabetically before choosing examples so identical subdomain
  collections yield deterministic statements and zero false interpretation
  changes during capsule comparisons. (#311)
• Include all tracked validation modules in source distributions: allowlist
  validation/run_path_safety.py, validation/evaluate_catalog_promotions.py,
  and four tracked validation notes in .gitignore so VCS-aware sdist builds
  package all files needed by tests. (#314)

Changed

• Refresh validation model helpers: default to current claude-sonnet-5-5,
  update pricing table and documentation examples with current model identifiers,
  and flag unknown model pricing instead of reporting a misleading $0.00. (#315)
• Support OpenRouter, Google Gemini, and local models (Ollama, vLLM, LM Studio)
  alongside Anthropic, OpenAI, and xAI in validation/agentic_ux and
  validation/triage_llm.py. (#316)