fix: include source titles in graph exports

[w3lld1]

What changed

• Adds source_title to graph JSON export nodes so graph consumers receive the same non-empty citation title fallback used by answer citations.
• Derives the title from the backing memory's section/header first, then the base source filename when source_section is empty.
• Adds regression coverage for subject graph JSON exports with an empty source_section.

Why this change exists

Graph export consumers could not rely on a human-readable source title when a memory had no source_section, even though citation titles already had a shared fallback path.

Touched surfaces

• cmd/cortex/main.go graph JSON export
• cmd/cortex/graph_subject_test.go

Tests

• go test ./cmd/cortex -run 'TestRunGraph_SubjectJSONIncludesSeedFactIDs|TestResolveGraphSeedFactIDsBySubject|TestRunGraph_SubjectAndFactIDMutuallyExclusive'
• go test ./...
• go vet ./...

Risk

Low. This adds an optional JSON field and leaves existing graph node fields unchanged.

AI-assisted implementation, reviewed locally with the test suite above.

Fixes #374

[hurttlocker]

Verified locally: clean build, and both graph subject tests pass (TestRunGraph_SubjectJSONIncludesSeedFactIDs, TestRunGraph_SubjectAndFactIDMutuallyExclusive). Reusing search.CitationTitleForResult instead of reimplementing the fallback is exactly the right call — one title path everywhere.

One non-blocking observation: graphSourceTitle does a GetMemory per node, and facts frequently share a backing memory, so large subgraph exports pay an N+1. Not worth holding this PR for — I'm filing it as a follow-up good-first-issue (memoize by MemoryID inside the export loop). If you want to grab that one too, it's yours.

This also closes the graph-export half of #374. Thanks for the clean regression test — merging.