Phase 4a: execution context — tests and the fake

[Wahbeh-Mohammad]

Part of #14. Second of three phase 4a PRs — the tests for #57. Targets that PR's branch; the phase record is #59. The same rebase-and-reprove note as #57 applies: the stack predates 4b's merge and rebases before it goes in.

What lands

12 files, +1,691: gems/dexpace-core/test/dexpace/** mirroring lib/ one file per file — the ten mirrors (context_test.rb, context_store_test.rb, error/context_conflict_error_test.rb, the three under context/, the four under instrumentation/), split into nested DexpaceTestCase classes per behaviour group for Metrics/ClassLength — plus the one double the design's testing strategy names and one gate test:

• FakeContext — include Dexpace::Context, a call_key and a store, nothing else. Every store rule (CTX-7–CTX-13, CTX-18, CTX-19) is about a keyed occupant, so driving them through a real chain would couple the store's suite to three other constructors; it gets #close free from the module, which is what makes the CTX-9/CTX-10 cases readable. Required by require_relative from exactly the two suites that use it, never from test_helper.rb.
• test/gates/rubocop_config_test.rb — pins the eighth cop's Include to every gem's lib/, because CopCase structurally cannot see .rubocop.yml and narrowing the scope to core alone survived the cop's own table.

Every file's header names the requirement IDs it exercises; every class inherits DexpaceTestCase; every thread is joined (the base class's teardown counts them); nothing uses assert_nothing_raised.

What the suites prove rather than restate:

• CTX-9's trap, three lines — two contexts with the same pinned call_key, so they are == and not equal? (the only pair that discriminates, R2): set the first, release the second — the slot is untouched and #release returned false; release the first — gone. Two default-constructed contexts would pass against an ==-based implementation.
• CTX-5/CTX-6/CTX-15, both halves — two DispatchContext.build(bundle: Bundle::NONE) not ==; the same pinned key ==, eql?, hash-equal; three flavours from one counter get three strictly increasing suffixes (strengthened after the per-flavour-counter mutation survived the weaker form); two keys from the same equal? Bundle::NONE differ.
• CTX-2 by identity — assert_same on the bundle and the key across both promotions, the source unchanged in the same test; assert_equal would pass against an implementation that rebuilt them.
• CTX-16 as a negative — a RequestContext promoted with and without an operation_name shares its key, its request and its store slot.
• CTX-8's race and CTX-7's burst — 32 threads released from one ::Thread::Queue barrier onto one key: exactly one winner, 31 ContextConflictErrors each naming the key; 16 threads × 1,000 keys: final size 16,000.
• CTX-11/CTX-12/XCUT-14's bound, not the vacuous count — from a store at cap, every further insert evicts exactly one and the size is never observed above cap; the aggregate inserts − cap would pass against a split lock and against no drain at all.
• CTX-13 through a real promotion — cap 3, chains a/b/c at request stage, a promoted to exchange, then d: a is gone, so re-setting a key through promotion does not refresh its position (the round-0 finding: the claim had rested on the fake only), and #close on the evicted context returns false and raises nothing.
• ContextStore.default in a fresh process — require "dexpace" alone leaves the singleton assigned before any call, so CTX-17's "construction registers nothing" stays inert by construction (the round-0 finding: a memoised .default had survived every suite).
• CTX-19's reachability — 1,000 registered contexts, locals dropped, three GC.starts, size still 1,000 and a sampled key resolving; a discriminator against ObjectSpace::WeakMap, and deliberately not against WeakKeyMap, which the cop covers.
• The Fiber[] boundary, with its guard — Fiber[:probe] visible and Thread.current[:probe] nil inside a child Fiber, a new ::Thread and an Enumerator's fiber (the round-0 finding: the guard had covered one carrier of the three), and the store finding the context in all three.
• CTX-20 — NO_TRACER_FACTORY.tracer from 16 threads, assert_same on all sixteen; the five-parameter call shapes.
• A Symbol or Integer call_key or operation_name refused on .build, #with and both promotions (the round-1 finding).
• Property tests through #sample: TraceIdFlavour.of round-trips; #renders?(x) == (#valid_trace_id?(x) || x == invalid_trace_id) over the three flavours — the two predicates disagree at exactly the sentinel, which is why both exist; Bundle.build(**bundle.to_h) == bundle.
• #with re-validation on the floor — on every one of the five Data types, since Data#with skips initialize on 3.2.11 and Model#with is what re-validates.

Verification

• bundle exec rake on Ruby 4.0.6 at this tip: exit 0, all seventeen gates green — test:gems 1,272 runs / 6,889 assertions across the six gems (118 of them the ten new suites), line coverage 99.96%; test:gates 130 runs; cops:test 129 cases; yard 402 methods, 0 undocumented.
• The matrix set on 3.2.11, 3.3.12 and 3.4.10, each with a fresh lock: 1,272 runs, 0 failures, 0 skips. Every interpreter-sensitive assertion — ObjectSpace::WeakKeyMap undefined on 3.2.11, Data#with skipping initialize there, the Fiber[] inheritance, the GC-based reachability — green on every row.
• The gem suite on four extra seeds (1, 99991, 424242 on 4.0.6; one non-default on 3.2.11): identical run counts, 0 failures; the two thread tests 15/15 under timeout.
• The three reviewers applied 18, 20 and 24 single-edit mutations themselves; every one caught at this tip on both interpreters except the plan's documented while→if drain equivalent. The two that survived round 0 — the release-then-set promotion and the memoised .default — are why the real-promotion CTX-13 case and the fresh-process case above exist.

Known follow-up from the review (not blocking)

• One report discrepancy, explained: the floor's tests-tip coverage reads 3,188 / 3,189 (99.96%) in the final review where a fix round reported 100.00%; the line is phase 2's registry.rb:253, a scheduling-dependent branch main's own suite covers or not run to run — not 4a code and not a 4a test.