articulate-writing 0.17.0

You can now see the rhetorical devices in a draft, hear the findings read
aloud, and see how much of each property an edit kept. The intensifier rule no
longer blocks learner writing under the default profile.

• filler-intensifier no longer blocks under the default profile. It still
  shows on every report, as a LOW finding labelled report only:. On the Liang
  et al. (2023) essays it blocked 26 of 91 TOEFL learner texts against 2 of 70
  length-matched US college essay windows. With the change, the TOEFL minus
  college blocked-rate gap on a held-out half goes from +0.24 to 0.00. A profile
  that wants the rule to block names it in gate_promote. The decision record
  is docs/calibration.md.

• A HIGH or MEDIUM finding inside a quotation attributed to someone else (a
  speech or citation verb beside the quote) moves to LOW with the label prefix
  quoted:. Quotation marks around your own paragraph change nothing.

• Both changes live in a new layer, articulate.calibration (calibration/1),
  after the pinned detector. The ruleset and its fingerprint are unchanged. A
  receipt names the layer when it moved a finding, and receipts made before it
  still verify.

• articulate devices reads thirteen rhetorical devices: antithesis,
  tricolon, anaphora, chiasmus, alliteration, assonance, polysyndeton,
  asyndeton, anadiplosis, epizeuxis, isocolon, zeugma and oxymoron. It reports
  where each one is and never gates. Rates on a held-out set written without
  sight of the readers are in docs/devices.md.

• --speak and --ssml on articulate devices, and --speak on
  articulate check, give a script of the findings to read aloud.

• The change report from --explain carries survival: a clarity value for
  each fact, device, rhythm, sentence order and register, 1 kept, below 1 lost.

• Reports and receipts now name their limit field scope, and each scope says
  what the result covers. The field was named does_not_prove before. These
  schemas move to version 2: changes, meaning, corpus-report, corpus-receipt,
  desk, interview, titles, restructure, voice-compare, house-receipt,
  process-summary, editor-receipt and the dialogue report. articulate verify
  still accepts version 1 corpus receipts, voice comparisons, house receipts and
  process summaries. To read either version in your own code, read scope and
  fall back to does_not_prove; articulate.scope.scope_of does this.

• Tool descriptions say "Runs on your machine". The findings tools say "Use the
  findings to improve writing. Use other evidence to decide who wrote a text."
  The CLI describes Articulate as a tool that finds writing patterns and shows
  where they are.

• articulate campaign reads a JSON list of short social texts and reports
  promotional templates in posts, stock-persona slot grammar in bios, and
  clusters of near-duplicate or templated posts with their shared skeleton.
  Each finding names the pattern, its offsets and its cost to a reader. The
  report format is articulate-campaign-report/1, with a JSON schema at
  docs/campaign-schema.json. Python API: articulate.campaign.analyze.

• A labelled synthetic corpus in corpus/campaign/ with ESL, dialect,
  translated, code-switched and non-English hard negatives, a per-platform
  split, a per-stratum fairness bound and controls that must fail. Run it with
  python -m articulate.campaign_bench. No change to the prose checker or its
  ruleset.