wiki-update: split tri-protocol ingest pipeline, integrate read-pdf fanout substrate

[nsmiller2501]

Summary

Follow-on to #8 (wiki-update). Restructures /wiki-update around three ingest protocols (E/M/S) selected automatically by substrate availability, hoists per-protocol prompts into sibling files, and extracts deterministic scripts.

⚠️ Depends on /read-pdf from PR #19 for the marker-fanout ingest path. Protocol M (the marker-conversion path) uses /read-pdf/scripts/prepare_substrate.py to generate the per-chunk substrate. PR #19 is itself a follow-on to #6 — when #6 and #19 land, this PR's Protocol M works. If only #6 lands without #19, Protocol M falls back to Protocol S (vision-batch) automatically, and the skill still functions. Conceptually I can't imagine accepting /wiki-update (#8) while rejecting /read-pdf (#6) since #8 already names read-pdf as a dependency.

What changed

Tri-protocol ingest

The skill now auto-detects which ingest path to use for each new PDF:

• protocol_e.md — "Existing extract." If a cached _text.md is already present (from a prior /read-pdf or /split-pdf run), use it directly. Cheapest path; zero PDF re-reading.
• protocol_m.md — "Marker." When /read-pdf's marker conversion is installed, route the paper through prepare_substrate.py for clean fanout reading. Highest fidelity for tables/equations/figures.
• protocol_s.md — "Split." Fallback when marker is unavailable; uses the legacy /split-pdf vision-batch path.

Sibling-file structure

• common.md: shared steps across all three protocols (first-run scaffold, citation-overlap check, project-lens summary, final /bib-update invocation).
• wiki_synthesis.md: project-lens summarization prompt and rules, loaded once per ingest.
• SKILL.md: now an orchestrator — detects substrate, picks the protocol, delegates to the chosen protocol_*.md, finishes with wiki_synthesis.md.

Scripts

• New scripts/scaffold_wiki.sh: shell scaffolder for references/raw/, references/wiki/, and references/CLAUDE.md on first invocation.
• New scripts/citation_overlap.py: detects when two papers cite the same prior work and flags it in the wiki cross-reference index.
• New scripts/copy_marker_figure.py: pulls figure images out of marker output and into references/wiki/<paper>/figures/.

skills/wiki-update/README.md updated to describe the tri-protocol ingest pipeline.

Why

• The original /wiki-update carried all three ingest paths in one SKILL.md, paying the full token cost on every invocation regardless of which path actually fired. The protocol split lets the orchestrator load only the relevant path.
• Tying Protocol M to /read-pdf's prepare_substrate.py unlocks marker-fidelity figure/table/equation extraction for the wiki workflow — the same upgrade /read-pdf itself gets in PR read-pdf: become canonical PDF-reading skill, absorb split-pdf as compat wrapper #19.
• The script extractions match the deterministic-script pattern shipped in PR bibcheck: split modes into sibling files and extract splitter script #13 (bibcheck), newproject: extract scaffold logic into scripts/scaffold.sh #14 (newproject), and bib-update: extract DOI/CrossRef/OpenAlex cascade into scripts/fetch_bibtex.py #16 (bib-update).

Testing

• Ran /wiki-update against a references/raw/ containing three new PDFs.
  • First PDF had no _text.md and marker was installed → Protocol M fired, ran prepare_substrate.py, produced a clean wiki summary with embedded figures.
  • Second PDF had a stale _text.md from an old /split-pdf run → Protocol E fired, reused the extract.
  • Third PDF was a scanned image-based PDF that marker couldn't handle → automatic fallback to Protocol S (vision-batch).
• Verified the citation-overlap script identifies shared citations across the three papers and emits the cross-reference table.
• Confirmed /bib-update is invoked as the final step and successfully appends new .bib entries.
• Verified first-run scaffold creates references/raw/, references/wiki/, and references/CLAUDE.md if absent.