Replace core parsing/resolution engine with Rubydex

[paracycle]

Summary

• Replace the entire AST walking, constant extraction, and resolution pipeline with Rubydex, a high-performance Ruby indexer written in Rust
• Rewrite association detection to use Prism's native AST directly, dropping the parser gem as a direct dependency
• Net result: +403 / -3,391 lines across 47 files

Architecture change

Before: For each file → parse with Prism → walk AST nodes → extract constant references via ConstNodeInspector → resolve via constant_resolver gem → check package violations

After: Index all workspace files in one Rubydex::Graph batch call → resolve all constants → iterate ResolvedConstantReference objects with direct links to target declarations → check package violations. Association detection (has_many, belongs_to, etc.) runs as a supplementary Prism-based pass since Rubydex doesn't understand ActiveRecord semantics.

Dependencies

Removed	Added
constant_resolver	rubydex
parallel
ast (direct)
parser (direct)

Kept: prism (for association detection), better_html (for ERB), activesupport, sorbet-runtime, zeitwerk, bundler.

Deleted files (16 source + 11 test + 3 RBI)

The entire per-file parsing pipeline: file_processor, node_processor, node_processor_factory, node_visitor, node_helpers, const_node_inspector, constant_name_inspector, constant_discovery, parsed_constant_definitions, reference_extractor, unresolved_reference, association_inspector, cache, parsers/ruby, parsers/factory, parsers/parser_interface.

Key design decisions

1. Full workspace indexing: index_and_resolve indexes ALL Ruby files in the workspace (not just the scoped check set) so Rubydex can resolve cross-package references. Only the scoped files are checked for violations.
2. Association pass: Kept as a separate Prism-native AST walk since Rubydex treats has_many :orders as a method call, not a constant reference. Uses graph.resolve_constant to resolve the inferred constant name.
3. ERB support: Parsers::Erb simplified to extract_ruby_source which feeds into graph.index_source.
4. No more caching: Rubydex's Rust engine is fast enough that the MD5-based file cache is unnecessary.
5. No more parallelism gem: Rubydex handles parallelism internally in Rust. The parallel flag is accepted but ignored.

Test results

All 136 tests pass (excluding spring_command_test and autoload_test which have pre-existing Ruby 4.0 incompatibilities unrelated to this change).

[exterm]

Cool stuff, Ufuk. Would this mean packwerk doesn't depend on the interrogated codebase using zeitwerk anymore?

[paracycle]

Cool stuff, Ufuk. Would this mean packwerk doesn't depend on the interrogated codebase using zeitwerk anymore?

That's correct. Rubydex can do proper Ruby constant resolution, so we don't need to use Zeitwerk heuristics to figure out what constant references resolve to based on their filename. At least, we shouldn't, and, if there are any problems with the resolution, then we should fix them in Rubydex.