Fix complexity blind spots, exact churn, a meaningful hot flag, and 2× faster runs

[giacope]

These fixes come out of running hashira on 15 MIT-licensed Rails apps and gems: rails, rubocop, chatwoot, feedbin, rubygems.org, huginn, errbit, fat_free_crm, postal, devise, activeadmin, good_job, sinatra, faker and hanami.

Fixes

Complexity only scored methods that sit directly in a class body (dc0b6ce)

• Before, methods written as private def / protected def, defined inside class << self or class_methods do, behind an if, or in a Const = Data.define do … end block were never scored.
• That also meant turning def into private def made a complexity regression disappear from --ratchet.
• A new Complexity::Body walk now covers the whole type body. It stops at nested classes and modules, so no method is counted twice. Singleton contexts name methods Class.method.
• Across the 15 repos this adds 24 complexity findings, for example faker's Faker::Internet.username (18).
• Heads-up: users' --ratchet will report these newly visible methods as NEW findings. This is noted in the CHANGELOG.

Churn matched files by bare suffix across the whole repo (08d83cf, 7dc824a)

• It took the maximum over every path ending in the file's name. Rails' action_view/base.rb was charged activerecord's 1218 commits; its own count is 352. A model picked up its spec's churn, and base.rb even matched database.rb. Only the first target's repo was read.
• GitLog now does one rev-parse --show-toplevel and one git log --full-history -- <targets> with literal pathspecs. It keys each path relative to its own target, so hits is an exact hash lookup.
• --full-history is needed because a path-limited log otherwise simplifies away side-branch commits.
• Checked against an exact per-path count: every hotspot row in all 15 repos now matches (0 mismatches). Before, 42 rails rows were inflated more than 2×.

"Both sites change often" was on 2265 of 2265 clones (c2d79dd)

• Any file with a single commit counted as changing often.
• A clone is now hot only when at least two distinct files it spans change more than the typical analyzed file, i.e. above the median commit count.
• Across the 15 repos this leaves 724 of 2265 clones hot, and same-file pairs no longer qualify.

Performance (byte-identical output)

• Duplication (16d5ebd): each statement's subtree is walked once per file through a shared Walks cache. Before, it was walked again for every window of up to 12 statements that held it.
• Smells (cbbd6ad): Lineage#assigned is memoized per class name. A namespace every file reopens, such as rubocop's RuboCop::Cop, used to redo the same ancestry walk hundreds of times.

full run	before	after
rubocop	48.6s	18.6s
rails (12 targets)	92.7s	42.7s
chatwoot	31.3s	16.1s

Across all 15 repos, every finding other than complexity and duplication is unchanged, and duplication clusters are identical apart from the hot flag.

Checks

• bin/ci passes: rubocop, rspec at 100% line and branch coverage, the self-gate, and the self-ratchet.
• kimera ci --since origin/main: 71 mutants on the changed lines, 0 survivors.
• Each memo has a spec that fails when the memo is removed. Two dead guards that mutation testing flagged were deleted.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CdcWkAxkqwRCup8VkCPcFA

---

Generated by Claude Code