v0.9.0

The 0.8.0 release build stopped before PyPI, so 0.9.0 is the first PyPI release
after 0.7.0 and carries everything listed under 0.8.0 as well. Detector rules
stay the published v0.5.2 rules (fingerprint sha256:9f78a7484bb20f84), and
with no project config and no domain profile, command output is byte-identical
to 0.8.0.

One release line. main now carries everything released from release/0.5.x
(0.5.1 to 0.8.0), so every later release comes from main.

• The writing-process record, disclosure statements and the review desk come
  over from main: articulate process, articulate disclose and
  articulate desk. fix and polish add an assistance entry to a document's
  process log when it has one. These commands add no detector rule and
  change no finding.
• Detector rules stay the published v0.5.2 rules (fingerprint
  sha256:9f78a7484bb20f84), so every 0.7.0 and 0.8.0 receipt still verifies.
  The reworked detector, the fairness harness and check --house-notes that sat
  on main change findings, so they stay out of this line until a reviewed
  ruleset change. Their history stays reachable from main.
• fix, polish and plan take --allow-change KINDS, and the MCP fix,
  polish and edit_plan tools take allow_change, so an edit can update a
  number, link, citation, quote or other named kind of protected content on
  purpose. Each allowed change is listed in allowed_changes in the result and
  the editor receipt. The plan binds the list, so a submission cannot widen it.
  Disclosure lines, added first-person sentences, HTML and math can never be
  allowed. Plans that allow nothing keep their old schema and plan IDs.
• fix, polish and submit take --explain [json|text], and the MCP fix,
  polish and edit_submit tools take explain, for a change report: each
  changed sentence before and after, the findings it cleared, kept or added,
  the paragraphs the guard kept and why, and each allowed change.
• Five domain profiles add rule packs for one kind of writing:
  ux-microcopy (UI string length, case, vague errors and link text),
  code-review (condescension, absolutes and requests with no reason),
  plain-language (reading grade, long sentences and wordy phrases),
  controlled-english (sentence length, one instruction per sentence, idioms,
  phrasal verbs and vague pronouns; no ASD-STE100 conformance is claimed) and
  rfc-keywords (normative-spec plus RFC 2119 and RFC 8174 keyword checks).
  Choose one with --profile, an in-file writing-profile: tag, the LSP
  server or an edit plan. The packs run after the detector, so the detector
  rules, every existing profile and every existing receipt are unchanged.
  Receipts under a domain profile record an extension_fingerprint, and a
  pack change makes them read Unverifiable. python -m articulate.bench also
  runs the new domain regression corpus.
• The hosted MCP service accepts edit plans built under a domain profile.
  It refused them before. A plan whose domain profile was altered is still
  refused.
• A project config file, .articulate.json, found by walking up from each
  file (nearest wins). It maps path globs to profiles, adds banned and
  preferred terminology with their own rule ids, allows terms of art, freezes
  terms that every rewrite must keep, and tunes rule-pack options. check,
  score, receipt, plan, judge, fix and polish take
  --config PATH|none, and articulate config PATH shows what applies to a
  file and why. Terminology findings appear in check text, JSON, SARIF,
  per-span verdicts, LSP diagnostics and receipts. Receipts embed the project
  rules with their hash, so they replay anywhere; an edited payload reads
  Unverifiable. A malformed config stops the command with exit 2, the file and
  the reason. With no config file, output is unchanged.
• The Windows x64 native package now carries the house voice spec. The 0.8.0
  release build stopped because the frozen server could not find that file at
  start-up, so 0.8.0 never reached PyPI. The release check now lists all 17
  local tools and runs house_transform in the frozen binary.