unslop

Docs and presentation release. No code changes; tests still 555 passed,
3 skipped.

Added

• docs/RESEARCH_AND_TECH.md — public reference compendium covering 38
  verified citations across 20 research categories, full tech-stack
  inventory, and design differentiators with file:line evidence.
• README Engineering & research section above Architecture: links to
  RESEARCH_AND_TECH.md, embedded research-depth constellation visual,
  inspirations table (blader/humanizer, Antislop ICLR 2026, Adversarial
  Paraphrasing NeurIPS 2025, DivEye TMLR 2026, Liang et al. 2023), and a
  "What we deliberately don't do" list.
• README Who this is for 3-persona strip (Engineers / Writers /
  Researchers).
• README author byline near the footer.
• Five new PNG visual assets: assets/hero-refresh.png,
  assets/research-depth.png, assets/social-preview.png,
  assets/demo.png, assets/statusline.png.

Changed

• README hero image swapped to hero-refresh.png; subtitle simplified.
• README jump-link bar reordered (Demo first, Research added) and TOC
  updated to match.
• README feature grid trimmed from 9 cells to 6; surprisal-variance,
  reasoning-trace sanitizer, and mode gating moved into a collapsed
  Power-user features details block.
• README voice-match feature card rewritten in plain English; dropped
  "stylometric" / "sycophancy-memory vector" jargon.
• README demo and statusline assets switched from SVG to PNG.

Fixed

• README: AI-ism reduction stat corrected from 89.1 % to 92.1 %
  (9-fixture suite, 2026-04-28 latest run).
• README: "Six modes, one command" → "Five modes, one command" (subtle,
  balanced, full, voice-match, anti-detector).
• README: skills table no longer claims /commit or /review triggers
  the commit/review sub-skills — only /unslop-commit and /unslop-review
  match the hook regex.
• skills/unslop/SKILL.md: subtle-mode description aligned with
  deterministic code path. Sycophancy and hedging stacks need at least
  balanced; subtle is stock-vocab-only.
• unslop/scripts/validate.py + soul.py: removed broken arXiv
  2604.11687 / "Kalemaj et al." citation. The ID resolves to an
  unrelated paper. Threshold kept; comment marked empirical pending
  re-verification.
• unslop/scripts/surprisal.py: added missing arXiv ID for the
  Chakraborty et al. ICML 2024 reference.
• tests/unslop/test_humanize.py: renamed shadowed TestHumanizeFileEx
  edge-case class to TestHumanizeFileExEdgeCases. The four main-path
  tests in the first class were silently dead due to Python class
  shadowing; they run again.
• 14 ruff F401/F811/F841 hits across tests/ and evals/measure.py
  cleaned up.

Removed

• requirements-optional.txt: dropped scipy (zero imports across the
  repo).
• 30+ stale benchmark result dumps (benchmarks/results/20260421T*.json,
  20260426T*.json, llm-mode-baseline.json) and the
  evals/snapshots/20260419T220738Z/ snapshot. Current Apr 28 runs and
  the baselines/ / humanness/ subtrees remain.
• Orphan SVG assets hero.svg, demo.svg, statusline.svg, and
  social-preview.svg. PNG replacements take over.