Speed up project effect inference - #45

Summary

• separate effect-only classification from provenance resolution so internal inference does not resolve an unknown reason for every call
• keep project modules in source-based fixed-point inference instead of falling back to compiled dependency BEAM analysis
• collect unknown dependency calls during the first fixed-point pass and preload each dependency module once with bounded parallel work
• preserve explicit provenance, direct dependency classification, typespec inference, and the dependency recursion guard

Addresses #44.

Performance

Measured on a 638-file Phoenix project with the same command: MIX_ENV=test mix reach.check --arch --smells

• Reach 2.8.2: 91.76s
• Reach 2.8.3: 473.98s
• this branch: 129.81s under concurrent system indexing load

Internal profiling showed the first effect-inference pass falling from 220.38s to 2.18s. Dependency preloading took 0.15s for 169 requested MFAs.

Verification

• mix ci
• 1,273 main tests and properties passed
• 23 calibration tests passed
• Dialyzer: 0 errors
• focused regression coverage verifies one BEAM load for multiple functions in one dependency module, no dependency fallback for project modules, and no module resolution during effect-only classification