fix: improve affected, callers, and context precision

[bohe76]

Summary

This PR combines three small CodeGraph precision improvements that were tested together:

• include semantic file dependents for affected by considering symbol-level edges and normalizing Windows-style paths
• include instantiates edges in caller lookup so class construction is reported as usage
• reduce noisy generic context entry points and keep output focused with fewer default code blocks plus compact same-file focus snippets

These changes are also available as split PRs for easier review:

• fix: include symbol references in file dependents #351: affected semantic dependents
• fix: include class instantiations in callers #355: class instantiation callers
• fix: reduce generic context entry noise #356: context noise/token reduction

Local evaluation

I ran the same 14-case local smoke suite against origin/main before these changes and against this combined branch.

Source	Accuracy	CodeGraph approx tokens	Savings vs manual
baseline f366222	6/14 (42.9%)	17,661	51.8%
this branch	14/14 (100.0%)	8,976	76.5%

Compared with baseline, the combined branch reduces CodeGraph output tokens by 8,685 tokens, or 49.2%, while passing 8 additional cases.

Token estimate method: ceil(character_count / 4) over prompt-visible command output. The smoke accuracy checks are string-presence checks for expected files/symbols, not a claim of semantic completeness.

Testing

• npm run build
• npx vitest run __tests__/graph.test.ts __tests__/context.test.ts --minWorkers=1 --maxWorkers=1 --testTimeout=30000 --hookTimeout=30000 --no-file-parallelism
  • 2 test files passed
  • 43 tests passed