[ffumero2003]

Problem

The exported helpers under x/ are used for fault testing, including by blazer's own integration tests. #176 lists five problems:

1. transport.Trigger writes the failure rate from a goroutine while RoundTrip reads it, which is a data race under -race. With context.Background() the goroutine never exits.
2. Synthetic responses (Response(status)) neither read nor close req.Body, and return a nil Header, which breaks the http.RoundTripper contract.
3. Stall still sends the request after the request context ends.
4. window.New divides by zero when size < resolution (on first use) or resolution == 0 (inside New).
5. internal/retry (used by every API call) and x/transport had no tests.

Fix

• x/transport:
  • The failure rate is stored atomically (atomic.Uint64 holding the float64 bits).
  • Trigger uses context.AfterFunc instead of a goroutine, so nothing outlives the context.
  • Synthetic responses close req.Body without draining it (draining could block forever on a pipe that is never written), and set Header, Proto and ContentLength.
  • Stall returns ctx.Err() when the context ends first, without calling the wrapped transport. It also closes the request body in that case, which Fix the x/transport data race and RoundTripper contract, the x/window divide-by-zero, and add tests #176 didn't ask for but the RoundTripper contract requires once the request isn't forwarded. Its godoc says so.
• x/window:
  • New panics with a documented message when a finite window has resolution <= 0, like time.NewTicker. No signature change.
  • It uses one bucket when size < resolution.
  • sweep compares absolute time-slot indices instead of array indices. The SDK panel found that with one bucket, the modulo comparison was always equal, so a window polled more often than its resolution never expired events.
• Tests: every x/transport option is covered (Trigger, Response/Body, Stall with a cancelled and a live context, MatchPathSubstring, FailureRate at 0 and 1, an unwritten io.Pipe body), in x/transport/failures_test.go, x/window/new_bounds_test.go and internal/retry/retry_test.go, in new files so they don't collide with test: add credential-free tests for calls only covered by live tests #231 and test: tighten minor sync, simulator and window test assertions #233.
• Changelog: two Fixed entries.

Verification

Each break was applied alone, failed its test, then passed after revert:

Break	Failing output
sweep back to the modulo comparison	Reduce after resolution boundary = 1, want nil
One-bucket rounding removed	integer divide by zero
resolution <= 0 check removed	New panic = runtime error: integer divide by zero, want documented resolution error
Trigger back to a goroutine	Trigger(context.Background()) added goroutines: before=2 after=34
Synthetic response leaves the body open	request body closed=false reads=0, want drained and closed
Synthetic response with nil Header	the synthetic-response assertion fails
Stall sends after the context ends	stalled RoundTrip error = <nil>, want context.Canceled
Backoff cap removed	Backoff(31s) = 1m3.3s, want capped below 32s
Backoff stops doubling	Backoff(1s) = 1.03s, want [2.02s, 2.06s)
Synthetic response drains the body again (io.Copy)	synthetic RoundTrip did not return while pipe body is unwritten
MatchPathSubstring matches every path	non-matching response = … StatusCode:500 …, want wrapped response
FailureRate(0) treated as 1	response = … StatusCode:500 …, want wrapped response
Stall with a live context skips the wrapped transport	wrapped RoundTripper was not called after the stall

With no B2 credentials set: go test -race -count=20 ./internal/retry/... ./x/..., go test -race -count=5 -shuffle=on ./..., go vet ./... and make ci pass. The live tests that use Stall (base/integration_test.go) still compile; their intent is unchanged, because the request was cancelled anyway.

Housekeeping in this PR

Two small changes outside #176's list, kept here because they touch the same files:

• 18558a7 ("chore: satisfy static analysis"): lint fixes the new tests needed.
• ioutil.NopCloser → io.NopCloser in x/transport/transport.go (io/ioutil is deprecated); no behavior change.

Deferred (pre-existing, not changed here): with an already-cancelled context, Trigger followed by a later FailureRate option leaves the final rate depending on option order.

Closes #176

[david-santacruz-backblaze]

Mate, please review following suggestions for this pr

• x/transport/transport.go, RoundTrip, o.status > 0 branch: io.Copy(io.Discard, req.Body) is unbounded and ignores req.Context(). Response(500) with an io.Pipe body that is never written or closed blocks past 2s. #176 item 2 covers pipe bodies. Close req.Body without draining, or drain in a goroutine that stops on ctx.Done().
• x/transport/failures_test.go: add a case with an io.Pipe body and a timeout. It should fail on the current code and pass after the fix.
• Split out the chore: satisfy static analysis commit and the ioutil.NopCloser to io.NopCloser swap, or note them in the PR description. Neither is in #176.
• Note in the PR description that Stall now closes req.Body on cancel. #176 only asks for ctx.Err().
• Change Window.New to window.New in the changelog fragment and state that it panics when resolution <= 0.
• Add tests for MatchPathSubstring, FailureRate at 0 and 1, and Stall with a live context. #176 asks for each option to be covered.

[ffumero2003]

@david-santacruz-backblaze Thanks, all six are addressed (c548485, c3e1103, 3071ff3, 4d054b8), and the branch is synced with main:

1. Unbounded drain: fixed. Synthetic responses now close req.Body without draining it.
2. Pipe test: added. TestSyntheticResponseClosesUnwrittenPipeBody uses an unwritten io.Pipe with a 2s bound; restoring the io.Copy fails it with synthetic RoundTrip did not return while pipe body is unwritten.
3. Out-of-scope changes: noted. The description has a "Housekeeping" section for 18558a7 and the ioutil → io swap.
4. Stall closing the body on cancel: noted in the description, with why (the request is not forwarded, so the contract requires closing it).
5. Changelog: fixed. The fragment says window.New and that it panics when a finite window's resolution is <= 0.
6. Option coverage: added. MatchPathSubstring (matching vs non-matching), FailureRate(0) and FailureRate(1), and Stall with a live context. Each has a break in the table that fails it.

go test -race -count=20 ./x/... ./internal/retry/..., go test -race -count=5 -shuffle=on ./..., go vet ./... and make ci (0 lint issues) pass.