Loki Mode

Fixed

• Proof-of-run (R1) now records the raw one-line brief from
  loki start "<brief>". Previously the inline brief was dropped and the proof
  page showed "No brief recorded" with source "codebase-analysis", even though
  the user gave a clear one-liner. The brief is written to .loki/state/brief.txt
  at brief-mode dispatch and surfaced verbatim in the proof (source "brief"),
  which is a stronger, more honest shareable artifact than the synthesized PRD.
  The marker is cleared at the start of every run so a later PRD or
  codebase-analysis run can never inherit a stale brief label. The brief still
  flows through the R1 redaction chokepoint before any published artifact, so a
  secret typed into a brief is redacted in the shared proof and gist.
• loki proof share gist URL no longer includes gh progress chatter: the
  uploader captured gh gist create stdout and stderr together, folding the
  multi-line progress text into the gist URL, the "Shared:" line, the
  ready-to-post hook, and LOKI_LAST_GIST_URL. stdout (the permalink) and
  stderr (progress) are now captured separately and the clean URL is extracted.
• loki wiki generate [path] accepts the documented positional path argument.
  The help text and design doc advertised it but argparse only defined --root,
  so a positional path errored with "unrecognized arguments". The positional now
  takes precedence over --root; --root and bare-cwd invocations are
  unchanged.
• Benchmark Claude Code adapter default model changed from the stale
  claude-sonnet-4 alias (HTTP 404 on current CLI) to the rolling sonnet
  alias. The adapter already parametrizes the model; this only fixes the default.

Internal

• local-ci hardening so the v7.16.1-class bug (a built dashboard SPA whose inline
  JavaScript threw a SyntaxError while still serving HTTP 200 and passing every
  presence check) cannot ship again: a new node-gated gate parses every inline
  <script> block in dashboard/static/index.html (compile-only, never
  executed; ES-module blocks are skipped to avoid false positives) and fails the
  build on any syntax error. tests/run-shellcheck.sh now prunes
  .claude/worktrees/ so transient agent worktrees do not pollute the lint set.
• This release is the close-out of a post-arc verification sweep: every shipped
  arc feature (proof-of-run, benchmark harness, cost dashboard, zero-config first
  run, auto-wiki + cited Q&A, 1-click rollback, open-core hooks, trust
  trajectory, team assets, marketplace) was re-tested the way a user uses it
  (fresh installs across npm/Docker/Homebrew, three-browser dashboard matrix,
  interactive Playwright click-through, a real free-backend end-to-end run, real
  gist/git-install/adapter checks, large-repo wiki). The four fixes above are the
  only product defects found; everything else verified clean.