no_human 0.1.1

Also in this release — reliability, honesty, and cost, measured not asserted
(full suite 8,864/0; funnel 5/5 with every holdout green; reviewer recall
17/19, up from 15/19):

• The eval judge's verdict now survives mid-run emission, a truncated end
  marker, and marker drift — six bench tasks per run were being scored as
  failures because a verdict could not be parsed, not because work was wrong.
• Git lock contention (another process briefly holding index.lock) is
  retried with two short backoffs instead of crashing the task; every other
  git failure still fails fast and loud.
• Fix pairs: when a task fails on an error this machine has overcome before,
  the retry is handed what worked — as evidence, never as an instruction.
• A retry that ends byte-identical to its predecessor (same failure, same
  diff) stops the loop and escalates honestly instead of buying the most
  expensive third attempt.
• Judgment-call blockers (ambiguity, novel-unknown, impossible) get exactly
  one supervisor-checked challenge before parking; external blockers are
  honored untouched, and a park is never converted into a fake "done".
• The reviewer carries a maintainability-trajectory lens: does this change
  make the NEXT change harder? Concrete findings only, capped below blocking
  severity.
• nh bench harvest: escalated, parked, and failed tasks become bench-spec
  candidates for curation.
• The intake grill's answering pass pays for what the task needs: probe
  budget scales with the question count; prose-only tasks skip filesystem
  probes (assumption-grade answers, clearly marked).
• Onboarding: two checkouts of the same repository are tellable apart —
  colliding names show their full path. (Authored end-to-end by no_human
  from its own board, review PASS, 8,847/0.)
• The stale-data banner no longer eats clicks while disconnected.
• docs: an operator profile for reviewing untrusted external PRs in a
  credential-isolated container.
• This release restores auto-update for installed apps: it ships the ZIP and
  latest-mac.yml that electron-updater requires (0.1.0's release lacked
  both).

Added

• CI builds the board-carrying wheel on every run and proves it installs:
  uv tool install <wheel> yields an nh that finds its board and the Agent
  SDK's bundled claude — no Node, no separate CLI install. A release build
  (workflow_dispatch with wheel_release) keeps the wheel as an artefact.
• A Claude Code plugin at plugins/no-human/ exposing the MCP bridge's two
  tools (task_add, task_status).
• A Publish to PyPI workflow (workflow_dispatch only, typed confirmation)
  that builds the board-carrying wheel and uploads it with PyPI Trusted
  Publishing — no API token anywhere in the repository.
• Version is 0.1.1 across pyproject.toml, desktop/package.json and
  web/package.json (and those lockfiles' root entries), so a built wheel is
  no longer labelled with the released 0.1.0's version.
• CHANGELOG.md (this file) and glama.json.

Changed

• README: download buttons, the site's hero loop under the title, install
  leads with the desktop app and names each build's architecture; nh approve
  is documented as what it does — it squash-lands the PR as the configured
  operator identity (git.approve_identity).
• CONTRIBUTING.md, docs/adapters.md and the nh task add --backend help no
  longer say "a single Claude backend": the coder runs on the Claude Agent SDK
  by default with OpenAI Codex as the sanctioned second backend
  (worker.backend); reviewer, planner, supervisor and utility tiers stay on
  Claude.

Fixed

• The shipped harvest test no longer asserts that the (unshipped) scored corpus
  directory exists, so the public repository's CI runs green.