wiki-compile

Run the paper-wiki wiki-compile action. This file is a Claude Code slash-command adapter; the action itself is platform-neutral.

Load the canonical project context

Read, in order:

1. WIKI.md — the only authoritative project rules, including variant, schemas, compilation rules, and prohibitions.
2. research.md — current scope, lifecycle, seeds/material inventory, and progress.
3. The existing wiki/ tree — current compiled state.

CLAUDE.md and AGENTS.md are runtime adapters. Never take schema or business rules from them, and never update them during compilation.

Diff raw/ against wiki/

List source material under raw/<topic>/, including OCR Markdown under raw/<topic>/mineru/. Match each source to its compiled note using the mapping in WIKI.md.

• New source: queue it.
• Already compiled: skip it unless the user requested recompile.
• OCR-derived source: validate paths and control metadata before reading any OCR Markdown. The source directory must be a direct child with a safe basename under raw/<topic>/mineru/; lstat it, its manifest, every listed Markdown file, and every ancestor from the wiki root. Each must be a real directory or regular file as appropriate, never a symlink, junction, hard link with an unexpected link count, or other reparse point, and every realpath must remain inside the source directory (the marker remains inside mineru/). Do not follow links while discovering files. "Safe" uses the OCR scripts' portable-name rules: no empty/dot component, separator, absolute/drive prefix, control character, trailing dot/space, Windows-forbidden punctuation or device basename, and no NFC+casefold collision.
• Only after those path checks, parse _paper-wiki-ocr-complete.json. Require schema: paper-wiki/ocr-completion/v2, state: requires-batch-commit, a 32-lowercase-hex batch_id, source exactly equal to the containing directory basename, and source_pdf to be a safe basename. Require content_fingerprint.schema: paper-wiki/ocr-content/v1, a nonempty files list covering every regular single-link file under the source except the completion manifest itself, and a tree_sha256. Each file record must contain exactly a safe relative POSIX path, nonnegative integer size, and lowercase SHA-256. Paths must be unique after NFC+casefold. markdown must be a nonempty, duplicate-free safe list exactly equal to the .md subset of files. Recompute the tree digest from UTF-8-byte-sorted file records using the canonical bytes paper-wiki/ocr-content/v1\n, then for each record path_utf8 + NUL + decimal_size + NUL + lowercase_sha256 + LF.
• Treat batch_commit_marker only as a basename matching _paper-wiki-ocr-batch-[0-9a-f]{32}.committed.json. Its named sibling under raw/<topic>/mineru/ must pass the same regular/non-link/non-reparse/containment checks. Require schema: paper-wiki/ocr-batch/v2, resolution: committed, the same batch_id, and exactly one source record matching all manifest provenance fields: source, source_pdf, source_pdf_size, source_pdf_sha256, and source_pdf_project_path, plus the complete content_fingerprint. If the project path is non-null, it must be a safe project-relative path inside raw/<topic>/ but outside mineru/; re-open that regular non-link PDF, verify its basename, size, and SHA-256 against the manifest, and block on any change. A null path preserves external flat-staging compatibility but must be reported as "current PDF hash not revalidated".
• Immediately before any OCR body is read, open each declared content file with no-follow semantics, bind its pre-open lstat, open-handle fstat, ancestor chain, and post-copy identity, and copy it into a new owner-only (0700 directory, regular private files) snapshot while hashing. From those snapshot bytes require the exact declared file set, every size/SHA-256, and the canonical tree_sha256. Read and compile only the verified private snapshot, never the original mutable path. Any legacy v1 manifest/batch or record missing the fingerprint is blocked; report that it needs re-OCR or an explicitly reviewed reseal operation.
• Queue the source only after every check passes. An unsafe source tree, invalid manifest, malformed/linked/mismatched/missing marker, unresolved pending batch, or any aborted marker for that batch_id is blocked and must be reported; never read its OCR body or compile it.
• Course source: apply WIKI.md and wiki/exam-scope.md; do not compile material that is out of scope.
• PDF without OCR/HTML/Markdown content: report that it needs OCR or a no-OCR source fetch before compilation.