Loki Mode v7.20.0

Changed

• Positioning refresh across README, SKILL.md, wiki, and the installation guide.
  Loki is now described as a spec-driven autonomous builder with a built-in
  trust layer (verified completion), surfacing the RARV-C loop, 11 quality
  gates, completion council, and the verified-completion evidence gate that were
  previously buried as internal mechanics. The wording matches the gate's actual
  behavior: an empty git diff against the run-start commit always blocks "done",
  and a red test run blocks completion when a test runner ran. No product
  behavior change.
• MCP server advertising corrected. The README and wiki previously claimed "15
  tools"; the server actually exposes 34 tools (26 in mcp/server.py, 7 magic
  tools, 1 managed-memory tool), plus 3 resources and 2 prompts. Numbers
  verified by counting registrations, not estimated.

Fixed

• Benchmark table honesty. The SWE-bench row claimed "299/300 patches", which
  read as a near-perfect resolution rate, but only patches were generated and
  the official SWE-bench evaluator was never run (no pass-rate result exists).
  It now reads "Not yet measured" with the exact reproduction command. The
  HumanEval row keeps 162/164 (98.78%) and now cites its results file
  (benchmarks/results/humaneval-loki-results.json) for provenance.

Notes

• AGENTS.md support (reading AGENTS.md with a CLAUDE.md fallback) was scoped for
  this release but deferred to a dedicated code release. Implementing it surfaced
  a separate, pre-existing dual-route parity drift in the no-PRD codebase-analysis
  prompt that must be reconciled first, rather than suppressed. Tracking it
  separately keeps this release a clean, parity-neutral positioning change.