v1.9.0

MINOR release. First phase of the synthesis output overhaul: paper.md
now compiles to a venue-correct PDF via Typst (10 venue templates),
the dashboard audit now checks CONTENT instead of just structure
(numeric grounding, accessibility, palette consistency, reviewer-skim
simulation), a cheap synthesis preview lets the researcher inspect
the predicted shape before drafting, and a content-depth audit BLOCKs
on stub-shaped IMRAD sections (no numbers in the abstract, no
statistics in the results, no limitations paragraph in the discussion,
no per-step coverage in methods). LaTeX path (tool_latex_compile)
is preserved for venues that require .tex submission.

Added — Typst PDF path

• src/research_os/tools/actions/synthesis/typst.py — md_to_typst,
  citations_md_to_hayagriva, compile_typst, paper_compile_typst.
• tool_paper_compile_typst — synthesis/paper.md → paper.typ →
  paper.pdf via Typst with venue-correct formatting. Returns
  pdf_path, page_count, citation_count, typst_warnings,
  parsed typst_errors with line numbers.
• 10 venue templates under src/research_os/data/typst/ (bundled
  in the wheel) and templates/typst/ (source checkout): nature,
  science, nejm, cell, ieee_conf, neurips, acl, plos,
  generic_two_column, generic_thesis. Per-venue citation style
  (Nature / IEEE / Vancouver / APA) wired through Hayagriva.
• writing_preferences.venue_template + writing_preferences.pdf_compile_engine
  added to the researcher_config.yaml template.
• [typst_compile] extra (no Python deps; opt-in marker).
• docs/VENUE_TEMPLATES.md documents every venue, the Markdown →
  Typst translation rules, and how to add a new template.

Added — Dashboard content gates

• src/research_os/tools/actions/audit/dashboard_content.py — seven
  sub-checks: numeric grounding (±1% tolerance vs workspace tables +
  citations), figure-to-text proximity, per-section substantiveness,
  WCAG 2.2 AA accessibility (contrast, alt text, heading hierarchy,
  button labels), print stylesheet sanity, color-palette consistency
  (Okabe-Ito / viridis / PuOr), reviewer-skim simulator.
• tool_audit_dashboard_content — wrapper; ungrounded numbers and
  missing-abstract-numbers are BLOCKERs, everything else WARNs.
  Supports override_dashboard_content_gate + override_rationale
  logged to workspace/logs/override_log.md.
• tool_dashboard_reviewer_sim — standalone "would a 5-min skim
  get the finding?" check.
• [print_audit] + [contrast_audit] opt-in extras.

Added — Synthesis preview

• src/research_os/tools/actions/synthesis/preview.py — cheap
  deterministic dry-run; reads the same sources as tool_synthesize
  but does NOT draft prose.
• tool_synthesis_preview(target, venue, mode) — returns
  predicted_word_count_per_section, predicted_total_word_count,
  predicted_page_count or slide_count,
  predicted_figures_embedded, predicted_citations,
  predicted_steps_drawn_from, detected_gaps,
  estimated_render_time_seconds. mode='diff' compares against
  the existing deliverable on disk.

Added — Content depth audits

• src/research_os/tools/actions/audit/content_depth.py —
  per-IMRAD-section audits beyond word counts: abstract needs ≥ 1
  number + ≥ 1 method + ≥ 1 conclusion verb, introduction needs
  ≥ 3 cited prior works + an explicit "in this study, we" pivot,
  methods must cover ≥ 80% of workspace steps (< 50% BLOCKs),
  results needs ≥ 1 statistic per primary finding, discussion needs
  a limitations paragraph + future-work direction, every cited key
  must appear in the bibliography.
• tool_section_substantiveness — runs all of the above; promotes
  to BLOCKER when ≥ 2 steps are uncovered in BOTH methods AND
  results.
• tool_audit_cliches — standalone scan for AI-cliché phrases
  ("in this study, we investigate", "future work should explore",
  "it is important to note that", etc.) with per-cliché replacement
  hints.

Added — reference projects + tests

• tests/fixtures/projects/paper_nature_minimal/ — exercises the
  full Typst PDF path under the Nature template.
• tests/fixtures/projects/paper_thin_content/ — same shape with
  intentionally stub-shaped IMRAD sections; verifies the content
  depth audit fires on cliché-shaped content.
• 77 new unit tests covering: every Markdown → Typst construct,
  Hayagriva citation conversion, end-to-end PDF compile for all 10
  venues (parametrised, gated on typst CLI availability), every
  dashboard content gate, synthesis preview predictions + diff
  mode, each per-section content depth audit, each cliché pattern.

Improved

• docs/TOOLS.md — added entries for the 6 new tools.
• tool.hatch.build.targets.wheel.force-include ships the bundled
  Typst templates inside the wheel.

Bumped

• 1.8.0 → 1.9.0 (pyproject.toml, __version__, CITATION.cff).
• Embeddings rebuilt to cover 207 tools + 150 protocols.

---