[travsr-analysis] #780 reconcile scip-ruby accessor/operator/DSL symbols

[Abhishek5517]

Issue

#780: on Ruby corpora, travsr status warned that ~52% of semantic definitions did not reconcile with the tree-sitter parse (fastlane @ 426146e: 8572/16527, 51.9%), so references could resolve to duplicate orphan nodes.

The 52% was three unrelated causes wearing one figure. This PR fixes all three, plus the tokenization and measurement issues that let them escape.

Result (fastlane, clean semantic reindex)

	attempted	missed	rate
before	16527	8572	51.9%
after	10697	36	0.34%

travsr fsck clean (no ghost nodes). Proof case supply/lib/supply/generated_universal_apk.rb: the 4 attr_accessor accessors (reader+writer) and == now have real method: twins, the scip: twins are gone, and travsr references package_name / travsr graph == resolve to the real node.

What changed

Prong A (measurement honesty + precision) - travsr-indexer, travsr-daemon, travsr-store:

• scip_name_kind parses the descriptor chain backtick-aware: the four metadata fields split on the 4th space (not split_whitespace().last()), and the leaf/container/paren splits skip separators inside backticks. RSpec DSL scopes whose descriptions contain spaces, # or / now parse intact instead of as garbage that still counted as a miss.
• is_synthetic_dsl_scope recognizes Sorbet DSL meta-scopes (describe/context/it blocks) that have no tree-sitter twin, distinguishing them from real <Class:>/<Module:> singleton scopes and from operator methods that start with <.
• unify_all drops synthetic DSL defs and defs in files tree-sitter never indexed (gitignored vendored code scip-ruby indexes anyway), excludes both from the attempted/missed counters, and drops their inbound refs/edges so they stop stealing spec-file references. Adds Store::phase_a_indexed_paths.
• A <Class:>/<Module:> leaf (a metaclass) unwraps to the real type so it unifies onto class:X.

Prong B (real reconciliation) - travsr-analysis:

• Capture setter (def x=) and operator (def ==, <=>, [], []=, ...) methods, whose name node is not an identifier.
• New post_parse hook on LanguageConfig (Ruby only) expands attr_accessor/attr_reader/attr_writer into reader/writer method nodes, and Struct.new into class:X plus member accessors, block methods, and the implicit initialize. Guards against double-emission when an explicit def also exists.

Prong C - travsr-analysis:

• Emit constants (const:X) and instance variables (field:C.@x) as def nodes so scip const/ivar references stop resolving to duplicates.

Notes

• Schema unchanged (v22, additive). No travsr-plugin-* src touched, so no plugin-hashes.lock regeneration.
• The other 9 config-driven languages get only post_parse: None.
• Unit tests added in ruby.rs, generic.rs path, indexer scip_unifier.rs, daemon scip_unifier.rs, and store lock each behavior in.

Out of scope (residual 36)

A handful of edge cases remain: methods added by reopening core classes (Hash), a few metaprogramming-defined methods, and classes defined inside describe blocks. These are distinct constructs, not part of this issue's root causes.

[github-actions]

Benchmark comparison (master → PR)

group             master                                 pr
-----             ------                                 --
bfs/chain/50      1.12   325.5±18.43µs        ? ?/sec    1.00    291.4±6.57µs        ? ?/sec
ppr/chain/100     1.14     32.6±0.71µs        ? ?/sec    1.00     28.5±0.21µs        ? ?/sec
ppr/chain/1000    1.12     32.7±0.38µs        ? ?/sec    1.00     29.3±0.19µs        ? ?/sec

Threshold: p95 regression ≥ +10% fails this check.