fix(test): remove the data race on attemptCount

[chengzeyi]

go test -race (the exact CI command) fails deterministically on main — 6/6 locally.

WARNING: DATA RACE
Read at 0x00c0004006a8 by goroutine 290:
  ...TestSubmitSingleShotOnConnectionFailure()  api/client_test.go:822
Previous write at 0x00c0004006a8 by goroutine 294:
  ...TestSubmitSingleShotOnConnectionFailure.func1()  api/client_test.go:802

The test goroutine reads attemptCount while the HTTP handler goroutine increments it. This test hijacks and closes the connection, so the client sees EOF and asserts while the handler is still unwinding — the other three counter tests happen to be ordered by the response write, which is why only this one trips the detector.

Changes (test file only)

• All four attemptCount counters → atomic.Int64. Fixing only the one that trips today would leave the same latent bug in the other three.
• The hijack handler called t.Fatal/t.Fatalf. testing.FailNow must only be called from the goroutine running the test — those are now t.Error/t.Errorf + return.

Verification

	Before	After
go test -race -run TestSubmitSingleShotOnConnectionFailure	6/6 fail	0/10 fail
go test -race ./...	FAIL	ok
gofmt -l	—	clean
go vet ./...	—	clean

No library code touched. atomic.Int64 needs Go 1.19; go.mod declares 1.20 and CI runs 1.22.

Not caused by the README merge — the same race is visible on the 2026-08-20 run of fix: align client behavior with the documented contract, which is where the test was introduced.