Key anonymous fixture caches on the declaration site

[ngan]

Anonymous fixtures — the inline fixture do … end / fixture extends: "x" do … end form — are cached on disk under a filename derived from the RSpec example-group constant name. That name is not a unique identifier for a declaration site, so two spec files can claim the same cache entry. When they do, the second fixture's generate is a no-op and its mount returns the first fixture's exposed records.

Why the identifier isn't unique

RSpec builds the group constant from the description via ExampleGroup.base_name_for, which strips every non-alphanumeric character, and resolves duplicates in disambiguate by appending a counter:

def self.disambiguate(name, const_scope)
  return name unless const_defined_on?(const_scope, name)
  name << "_2"
  name.next! while const_defined_on?(const_scope, name)
  name
end

Two problems, and the second is the one that bites:

(a) base_name_for is lossy. Foo::Bar, "Foo Bar" and "Foo-Bar" all collapse to FooBar.

(b) The disambiguation counter resets mid-process. It only appends a suffix while the earlier constant still exists, and RSpec::Core::World#reset starts with RSpec::ExampleGroups.remove_all_constants. Distributed and batched test runners call world.reset before loading each batch, so the first group of a given description in every batch gets the unsuffixed name.

Demonstrated with nothing but rspec-core:

--- both groups declared in the SAME batch ---
group 1 -> some_description
group 2 -> some_description_2   # RSpec disambiguated
same cache file? false

--- each group declared in its OWN batch (world.reset between) ---
batch 0 -> some_description
batch 1 -> some_description     # no suffix this time
same cache file? true

Why that becomes wrong data

The cache directory is cleared once per process (Runner#start, from before(:suite)), and Fixture#finish drops only the in-memory copy for anonymous fixtures — the JSON file outlives the batch that wrote it. So Cache#exists? reports another declaration's file as "already generated" and Fixture#generate returns early:

def generate(force: false)
  return if @cache.exists? && !force

Cache#load then builds the repository from that file, and Repository defines its readers from whatever names it finds:

NoMethodError: undefined method '<name>' for #<FixtureKit::Repository:0x…>

The worse case is when it doesn't raise. If both definitions happen to expose the same names, the readers exist, the suite passes — against the wrong records. That silent case is the real reason to fix this rather than work around the flake.

The fix

Fold a digest of the declaration site into the anonymous identifier:

_anonymous/foo/with_fixture_kit/hello.3f2a9c1d4b8e.json

Definition#fingerprint is the block's source_location plus its extends: target. source_location uniquely identifies each fixture do … end, is stable across processes, and doesn't depend on the test framework's constant naming at all — the information was already there, just unused. extends is included because a helper that declares fixture(extends: some_variable) gives every caller the same file and line, and sharing a cache entry there would mount the wrong parent's records.

The readable slug stays as a prefix so cache files and on_cache_saved logs remain browsable. The digest is joined with a dot rather than an underscore: scope names are snake_case, so an underscore leaves no parseable boundary, while a dot cannot appear in a slug (RSpec strips non-alphanumerics from the description before it is underscored, so even "a.b.c" becomes abc).

identifier_for is untouched, so adapter subclasses keep working, and MinitestAdapter is covered too since the digest is applied in Cache rather than in either adapter.

Also adds Registry#claim_cache_identifier, called from Fixture#generate, so any future collision raises and names both declaration sites rather than silently mounting the wrong records:

FixtureKit::CacheIdentifierCollision:
  fixtures declared at …/batch_three_collision.rb:10 and …/batch_four_collision.rb:7
  both resolve to the cache identifier '_anonymous/queue_mode_colliding_description'

It deliberately allows two fixtures built from the same declaration to share an entry — that's the shared-example-group case, where two host groups legitimately share one definition and raising would break working suites.

Testing

The failure is reproduced as a regression test rather than asserted: the existing queue-mode simulation gets two more batches whose top-level descriptions collide, each declaring a different anonymous fixture. Against the unfixed gem that produced the NoMethodError above; it now passes.

I verified the test actually gates on the fix by reverting only the identifier change — with the digest gone, the new CacheIdentifierCollision fires instead, which also confirms the backstop works end to end.

Unit coverage added for Definition#location / #fingerprint, for anonymous identifiers whose scopes stringify identically, for definitions extending different parents, and for the registry guard.

209 examples, 0 failures, on both the rspec and minitest integration paths.

Upgrade note

Cache naming changes, so existing tmp/cache/fixture_kit contents are ignored. Harmless — they regenerate — but expect a one-time cost on the first run after the bump.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XgYUaNMUsHSWNn1HEwqgrT