Add a CellStructureRecord

This skill turns one named, in-scope microbial structure into a validated
CellStructureMech record.

Use discover-cell-structures when the task is to search broadly for missing
structures or triage a literature lead. Use literature-evidence when the task
is to audit readability, propose verbatim snippets, or verify that a snippet is
actually present in a cited paper. Use source-queue for database or atlas
adoption.

Read first

• CLAUDE.md for branching, mutation, history, and PR rules.
• docs/CURATION.md for record boundaries, identifiers, evidence, components,
  traits, causal graphs, and images.
• docs/SCHEMA.md for the available CellStructureRecord sections.
• just report for live corpus counts.
• The closest existing data/structures/<category>/*.yaml record, to copy
  local shape but not facts.

Accept or reject

Add the structure only if it is a named subcellular entity in microbes:

• organelle, envelope layer, appendage, microcompartment, inclusion,
  cytoskeletal system, or multi-protein complex
• finest record no smaller than a multi-protein complex, unless a
  homo-oligomer forms a visually and biologically distinct structure
• at least one literature-backed component class
• at least one literature-backed function or assembly mechanism
• at least one real organism used as a canonical example

Reject single proteins, bare genes, pathways, phenotypes, broad placeholder
parents, storage compounds without an organized boundary, and host structures.
Record the reason in notes or in the response so the same near miss is not
re-triaged as an unknown.

Prove it is new

Before writing a record, search exact identifiers and names across the whole
repository, including ignored and hidden files:

rg --no-ignore --hidden -n \
  "<GO CURIE>|<minted slug>|<label>|<synonym>|<core component>|<DOI>|<PMID>" \
  .

Search identifier, xrefs, synonyms, has_part, part_of,
parent_structures, causal_graphs, discussions, generated pages, history,
and research/. If a prior mention is only a rejection, read it before
continuing. Never search broad prefixes such as DOI:10 or GO: to prove
absence.

Identity

Prefer an exact GO cellular-component CURIE.

1. Check the GO label, synonyms, definition, and parents.
2. Use the GO CURIE only when it denotes exactly the structure being recorded.
3. If GO is broader than the structure, mint cellstructuremech:<slug> and put
   the broader GO term in parent_structures only when it is a true is-a
   parent.
4. Use part_of when the new record is a layer, subassembly, or substructure
   of an enclosing structure.

Never guess CURIEs. If a candidate GO, InterPro, Pfam, CHEBI, SO,
ComplexPortal, UniProtKB, or NCBITaxon accession has not been resolved at its
issuing authority, leave it out and add a CURATION_TODO discussion that says
what must be checked.

Evidence bundle

Make the first record small but complete enough to stand alone:

• definition: GO for exact GO-grounded records, otherwise the paper that
  defines the structure
• components: taxon-agnostic protein, RNA, lipid, polysaccharide, or
  peptidoglycan classes; organism-specific UniProt accessions only in
  protein_examples
• taxonomic_distribution: clades and presence levels backed by a source,
  never inferred from one imaged strain
• canonical_examples: organisms actually studied in the cited papers
• functions: biological outputs, GO-grounded when an exact process or
  molecular function exists
• causal_graphs: at least one MECHANISTIC ASSEMBLY or FUNCTION graph
  with a citation on every edge
• evidence: record-level DOI or PMID references for the major review and
  canary primary papers

Use a review for broad scope and primary papers for at least one component,
canonical organism, or causal edge. Put only exact copied text in snippet;
put paraphrase and interpretation in notes.

Causal graph

The first causal graph should be readable and source-bounded:

• 5 to 8 nodes for a known mechanism
• at least 5 directed edges when the mechanism supports them
• component_ref on every node that is one of the record's components
• final structure node grounded to the record identifier when the record has
  an exact ontology CURIE
• subject, predicate, object, description, and evidence on every
  edge
• predicate_id only when an exact relation CURIE has been curated

Generic states, capacities, and intermediates may stay ungrounded. Do not add
an edge just to make the graph look complete.

Write the record

Create the skeleton with the validated scaffolder:

just new-record \
  --category <CATEGORY> \
  --kind <STRUCTURE_KIND> \
  --identifier <GO-or-cellstructuremech-CURIE> \
  --label "<label>" \
  --definition "<one-sentence definition>" \
  --definition-source <DOI-or-PMID-or-GO-CURIE> \
  --curator claude \
  --llm-assisted \
  --apply

Then fill the rest by loading the YAML with yaml.safe_load, appending a
record_curation_event(..., llm_assisted=True), and writing it with
write_validated_structure. Do not hand-serialize YAML or loosen the
round-trip test if formatting drifts.

Set mapping_status: PROPOSED for every model-drafted record. Promotion to
REVIEWED is a human decision.

For every added record, create repository-level history:

just new-history \
  --kind record \
  --slug <slug> \
  --target-root data/structures/<category> \
  --event CREATE \
  --outcome changed \
  --sections identity,composition,taxonomy,functions,causal_graphs \
  --summary "<short summary>" \
  --details "<what was added and which sources justify it>" \
  --actor-name claude \
  --model <model> \
  --agent-tool claude-code

Validate

Validate the new record directly with LinkML before broader gates:

uv run linkml-validate \
  -s src/cellstructuremech/schema/cellstructuremech.yaml \
  --target-class CellStructureRecord \
  data/structures/<category>/<slug>.yaml

Then run the checks whose scope LinkML does not cover:

just validate-strict --quiet data/structures/<category>/<slug>.yaml
just validate-history history/records/<slug>
just check-curies-strict
just validate-products
just render
just docs-stats
git diff --check
just qc

Also run just evidence-verify before committing a new snippet, and run
just check-trait-links --check after adding or editing associated_traits.

Report

End with:

• new identifier, label, and file path
• strongest identity source and strongest mechanism source
• every DOI, PMID, CURIE, source accession, and taxon id added
• history record and generated pages
• validation commands that passed
• any CURATION_TODO left open
• whether duplicate/absence searches included ignored and hidden files