tyler_med v0.1.2

Changelog

v0.1.2

Data-integrity fix for adding folders to an existing wiki.

• The index is now additive. Previously, running a conversion on one folder
  rebuilt index.md/index.csv/index.json from only that folder's papers,
  silently dropping every other paper from the index (their papers/*.md stayed on
  disk). The index is now rebuilt from all cached metadata — the whole wiki — so
  adding a folder unions into the existing index. --force still wipes the cache for
  a clean single-folder rebuild.
• Mount-stable cache keys. State is keyed on each paper's output filename instead
  of its absolute path, so a prior session's cache still matches when the workspace
  re-mounts under a new root (fixes cross-session incremental resume and the recursive
  rebuild). Legacy absolute-path caches are migrated automatically on load.
• --index-only needs no PDFs. The recovery path now rebuilds the full index from
  .wiki_state.json alone, without enumerating any PDFs — so it works even pointed at
  an empty/foreign folder.
• Guardrail. If the index ends up with fewer entries than the number of paper
  files on disk, the script prints a loud warning instead of shipping a shrunken index.

v0.1.1

Robustness and metadata-quality release.

Resumability / large or time-capped runs

• State is saved after every file; the index/evidence table is rebuilt from
  state each pass (--index-only to rebuild without converting).
• --time-budget SECONDS for time-capped, no-background sandboxes (Cowork): re-run
  the same command until it prints ALL_DONE. Already-converted files are skipped
  cheaply (size + mtime, no re-hashing) so passes converge; work is ordered
  smallest-first so one big file can't starve the queue; a guard warns if --force
  is combined with --time-budget.
• --max-pages N page-caps oversized documents (records the truncation in
  frontmatter) so a single huge PDF fits within one call.

OCR

• --ocr {auto,off,force} (default auto): skips OCR for born-digital journal PDFs
  (~4–5× faster per file, identical extraction) and OCRs only true scans.

Metadata quality

• Repair the U+FFFD (�) glyph collapse in high-confidence statistical contexts
  (e.g. N � → N =, � 80% → ≥ 80%, mean �SD → mean ± SD), which also
  recovers many previously-blank sample sizes. Ambiguous cases are left untouched.
• Cleaner author extraction: prefer the embedded PDF author when it reads like
  names, reject affiliation/address lines, capture multi-line bylines, strip orphan
  superscript markers.
• Controlled Obsidian tags — only design/…, source/…, and year/…
  namespaces (per-keyword hashtags moved behind --keyword-tags, off by default);
  journal boilerplate ("Article history: Received …") no longer bleeds into keywords.
• Broader sample-size detection (more phrasings and noun forms).

Docs

• SKILL.md: a repeatable index_by_theme.md (by-topic) routine; the conversion
  command now references the skill's own base directory, so it resolves in packaged
  or Cowork deploys where the folder is named tyler-med.
• README.md: an Obsidian + Dataview usage section.

v0.1.0

Initial release: a two-tier Markdown wiki (study-design-grouped index.md + full
papers/*.md) plus a index.csv/index.json evidence table, with reliable titles
from embedded PDF metadata + DOI, study-design classification, sample size, data
source, trial/PROSPERO registration IDs, mojibake/ligature repair, and duplicate
detection. MIT-licensed; derived from the tyler skill in econtools.