feat: graph and backlinks parse frontmatter links field

[akeemjenkins]

Problem

The okf graph and okf backlinks commands only parse markdown body links. The frontmatter links: field (a YAML list of concept IDs) is completely ignored. An LLM authoring concepts naturally puts links in frontmatter, but the graph shows 0 edges.

Fix

• Add Links []string field to concept.Frontmatter struct
• Add ExtractFrontmatterLinks() in validate package
• graph.Build() and backlinks.Backlinks() now parse both body and frontmatter links, deduplicated via existing edgeSet

[akeemjenkins]

Review — reviewed against current main (d641562), test-merged locally

Verified: merges cleanly (including alongside main's graph.go lint fix); go build / go vet / go test ./... pass, and golangci-lint is clean (0 issues) on the merged result.

Nicely factored. ExtractFrontmatterLinks returns the same Link struct, so frontmatter targets flow through the identical resolveLink path as inline links (relative/absolute/.md handling stays consistent for free), and dedup between body and frontmatter links is correct in both graph (edgeSet) and backlinks (seen + break), with tests proving the overlap collapses.

⚠️ Important — a single-string links value crashes the whole bundle load

Links []string (concept.go:25) only accepts a YAML list of strings. Verified empirically:

links: /b             -> cannot unmarshal !!str into []string
links: [/b, /c]       -> ok
links: [{target: /b}] -> cannot unmarshal !!map into string

Because bundle.Load propagates the parse error, one concept writing the very natural links: /b fails the entire bundle, not just that file. Suggest a custom UnmarshalYAML that accepts scalar-or-sequence (or, at minimum, don't let a malformed links take down the whole load, and document the expected shape).

⚠️ Important — the validator ignores frontmatter links

validateLinks still only scans c.Body, so a frontmatter link pointing at a nonexistent concept is silently dropped by the graph and never reported by okf validate. This PR is what creates the divergence (graph/backlinks now consume frontmatter links; validate doesn't), so it's the right place to also iterate ExtractFrontmatterLinks(c.Frontmatter.Links) in validation.

Nits

• Self-links: a concept listing itself yields an a -> a graph self-edge (backlinks guards this via c.ID == conceptID; graph does not).
• append(bodyLinks, fmLinks...) aliases the returned slice — harmless today, but worth copying to avoid a future footgun.
• Coverage gaps: duplicate frontmatter links, an FM link to a nonexistent concept, self-links, and the malformed-shape cases above.

Recommendation: approve after making links parsing robust (at least not fataling the whole bundle). The validator-consistency fix is strongly encouraged in this same PR since it's the source of the inconsistency.

_{🤖 Automated review via Claude Code, verified locally against current main.}