Phase 5a: configuration and the clock — tests and doubles

[Wahbeh-Mohammad]

Part of #18. Second PR of the nine-PR phase-5 stack — the tests for phase 5a's code PR below it. Targets that PR's branch; the phase record is the next PR up.

What lands

22 files, +3,286: nineteen suites under gems/dexpace-core/test/dexpace/ mirroring lib/ one file per file (the three private_constants — ConfigParsers, ProxyResolution, DeepValue — are asserted at their call sites and say so), plus the doubles the design's testing strategy names, one per file under test/support/, and a matrix-fact suite that pins the six floor-straddling facts on every row:

• FakeClock — a settable monotonic instant the test advances, a recorded sleep, exactly the three _Clock operations; a FakeClock-driven Clock#sleep and a deadline expiry need no wall time.
• FakeConfigSource — the hermetic env / property seam (CFG-11); renamed from the plan's FakeSource, which is phase 3a's IO-17 double at the same path.
• ParkingScheduler — the minimal Fiber::Scheduler CFG-18 needs: parks a fiber with a deadline, runs its loop in #close; a new file, because phase 2's ProbeScheduler at the plan's path is a hook recorder whose #block cannot drive a timed pop. All three top-level, the convention of every existing double (P5-58).

Every file's header names the requirement IDs it exercises; every class inherits DexpaceTestCase; every thread is joined; nothing uses assert_nothing_raised; no 5a case reads the real process environment — every source is injected (CFG-11), the slot's env seam is a FakeConfigSource wherever the slot is consulted, and the one case that could not be (the ENVIRONMENT seam itself) writes and restores a uniquely named key. Review round 0's blocking finding was exactly this class — the CFG-28 case resolved through the empty slot's real ENV and failed with HTTPS_PROXY set — and the fix commit made the proxy, ceiling and cap suites hermetic and reproduced each red first.

What the suites prove rather than restate:

• CFG-1–CFG-4 with every tier populated for one key and removed one at a time; the CFG-2 asymmetry (an empty environment value falls through, an empty override or property resolves to "") in one case so emptiness cannot become chain-wide; #string finding dexpace.max.retries from DEXPACE_MAX_RETRIES and #raw_property not.
• CFG-13 — 16 threads configuring and reading concurrently with no torn or nil observation; the slot's two publications pinned inside the mutex by text, because the race is unobservable under the GVL.
• CFG-15/CFG-17 — a 5 s sleep cancelled from another thread at 50 ms returning in 0.050 s with the token cancelled; an already-cancelled token returning at once; the wait re-checking the token after the wake (the mutant that drops it goes red on both interpreters).
• The bounded wait — Future#value(deadline:, clock: fake) on an unsettled Completer raising CancelledError with reason :deadline_expired and the completer settled cancelled; Completer#await still returning self; deadline: nil behaving exactly as phase 2 (its suite unchanged); a settled future still refusing an invalid deadline (from the review).
• CFG-18 — under ParkingScheduler, #block fires once and #kernel_sleep never; without a scheduler, SeamError naming Fiber.set_scheduler.
• The proxy model — '*.example.com' matching A.Example.COM and not example.com; CFG-26's a\|b|c → ['a|b', 'c']; port 65536, garbage and a port-less URL each yielding nil plus exactly one WarningCaptured warning, never a raise; the compiled pattern assert_same across lookups; a blank HTTPS_PROXY from any tier no longer masking HTTP_PROXY; the explicit no-argument .resolve reading the process-wide slot through fake seams installed by Dexpace.configure.
• CFG-22 — both credentials masked in #to_s and #inspect, with the username-only and password-only shapes (review round 0's should-fix).
• HTTPDate — GMT/UTC/+0000/+00:00 parsing to one instant; a wrong weekday and a lower-case month tolerated; '', ' ' and a missing comma refused; 31 Nov, 29 Feb 1995, hour 24 and second 60 refused (P5-54); English month names under a localedef'd de_DE.UTF-8.
• UUID — 10,000 distinct v4 values; 8 threads × 1,000; three distinct Random objects across the main fiber, a child fiber and a thread; a text scan of uuid.rb's non-comment lines for SecureRandom (because securerandom is allowlisted, the require gate cannot catch it).
• DeepValue through Configuration — NaN equal and hash-equal with two distinct NaN objects (the identity short-circuit trap), {a: 0.0} != {a: -0.0}, [1] != [1.0] (P5-14), a cycle returning; the Float#hash disagreement between the two NaN payloads asserted as the precondition.
• Retryability over the whole set, a Status object and an Integer; respond_to? false for any throwable-half name.
• The two wirings — ContextStore.default absent in a fresh process before the first call, sixteen first callers under a 50 ms seam getting one store (a slowed new, since the seams now run outside the lock), no seam consulted after the first call; IO.max_materialized_bytes at the default and configured, the four readers going through it (a reader put back on the constant goes red).
• The matrix facts, pinned on every row — Time.httpdate's tolerances; Random::DEFAULT gone; Queue#pop(timeout: -1) and a closed queue; RbConfig absent under --disable-gems; a Data keeping an ivar set before super through the freeze; Float#hash on the two NaN payloads.

Verification

• bundle exec rake on Ruby 4.0.6 at this tip: exit 0, all seventeen gates green — test:gems 1,762 runs / 44,284 assertions, line coverage 99.97% (the one line is phase 2's registry race branch); honest RuboCop 350 files clean.
• The matrix set on 3.2.11, 3.3.12 and 3.4.10 with a fresh lock: 0 failures; every interpreter-sensitive assertion (#with on every 5a Data re-validating on 3.2.11; the carriers; Queue#pop(timeout:); the scheduler; NaN/signed-zero hashing) green on every row.
• Four extra seeds on two interpreters: identical run counts; the six hang-shaped suites stable across five repeats each. The two reviewers applied 47 and 53 single-edit mutations; every one caught, ten on 3.2.11 as well.
• Hermeticity reproduced by the final reviewer: the proxy suite green with HTTPS_PROXY/HTTP_PROXY/NO_PROXY exported and with a port-less HTTPS_PROXY; the ceiling and cap suites green under hostile MAX_MATERIALIZED_BYTES/MAX_TRACKED_CONTEXTS.

Known follow-up (not blocking)

• R1-1 — builder_test.rb:203: Configuration::Builder.new(property_source: nil).build.string("K") builds over Sources::ENVIRONMENT and reads the host's K (K=hostile makes it fail). Pass env_source: Sources::NONE; the case is about property_source: nil.
• Phase 3a's and 3b's materialisation suites now read the ceiling through the slot with no fake seam (they assumed the constant before); stated in io_ceiling_test.rb's header rather than fixed here.