fix(providers): retry connect timeouts and proxy CONNECT failures

Three classification gaps against #447, all raised in review.

A connect-phase timeout was never retried. It is provably pre-send for the same
structural reason a refused dial is, since the connection was never established
and no request bytes left this host, but it arrives carrying "i/o timeout" and
the blanket wording exclusion rejected it before the connect gate could admit
it. Intermittent connect timeouts are exactly the transient failure the issue
is about, so the most common case was the one not covered. It is now admitted
ahead of the wording exclusions, gated on the connect phase so a post-send read
timeout, which is also Timeout(), still cannot replay a POST.

A failed CONNECT through an HTTP proxy was never retried either. net tags it Op
"proxyconnect" rather than "dial", so it missed the gate entirely. Same
guarantee applies: the tunnel to the model host was never built. The gate now
accepts both, which matters most in the environment where transient connect
failures are most common.

ETIMEDOUT and WSAETIMEDOUT join the errno lists so a connect timeout still
classifies if it reaches us with the errno but without the timeout flag.

The pre-send budget also now resets after a hop that reaches the server. It is
meant to bound consecutive connect failures, not the lifetime of the call, and
without the reset an early run of blips permanently spent it, so a later blip
surfaced as a transport error while the status budget still had room. That is
the coupling the two counters were split apart to remove, in the one ordering
the earlier tests missed. Total work stays bounded, since every reset costs a
status attempt and those are capped.

Also corrects the SendWithAuthRetry comment, which still claimed the inner path
retried only 429/503/529.

Tests cover dial and proxyconnect timeouts as retryable and read/write timeouts
as not, a refused proxy CONNECT, and the interleaved pre-send/status/pre-send
sequence. Each fix fails its test when reverted.

internal/providers/providerio/auth.go

@@ -25,7 +25,9 @@ type TokenResolver func(ctx context.Context, forceRefresh bool) (header string,
 //
 // base carries the API-key auth config (key + default header/scheme). setExtra
 // (optional) sets the provider's non-auth headers on every attempt. Transient
-// retries (429/503/529) are handled by the inner SendWithRetry.
+// retries are handled by the inner SendWithRetry: 429/503/529 on their own
+// budget, and provably pre-send transport failures (refused or timed-out
+// connect, DNS, TLS handshake) on a separate, shorter one.
 func SendWithAuthRetry(
 	ctx context.Context,
 	client *http.Client,

internal/providers/providerio/retry.go

@@ -3,6 +3,9 @@ package providerio
 import (
 	"bytes"
 	"context"
+	"errors"
+	"io"
+	"net"
 	"net/http"
 	"strconv"
 	"strings"
@@ -17,13 +20,22 @@ import (
 // (previously only the OpenAI provider retried; Anthropic and Gemini surfaced the
 // first failure).
 //
-// What is retried: ONLY 429 (rate limit) and 503 (service unavailable) — the
-// statuses where the server explicitly did NOT accept the request, so replaying
-// it cannot duplicate work. Other 5xx (500/502/504) and transport/network errors
-// are NOT retried: a completion POST is non-idempotent and may have reached and
-// been processed by the server, so replaying it could duplicate
-// (billable) work. Only the INITIAL request is ever in scope; once the response
-// body starts streaming it is never re-issued.
+// What is retried: 429 (rate limit) and 503 (service unavailable) — the statuses
+// where the server explicitly did NOT accept the request — plus PROVABLY
+// pre-send transport failures (DNS resolution, a refused/unreachable dial, a TLS
+// handshake timeout) where no request bytes ever left this host. In every one of
+// those cases we KNOW the server did not receive and process the request, so a
+// replay cannot duplicate work. Other 5xx (500/502/504) and every ambiguous or
+// post-send transport error (connection reset, broken pipe, EOF, a generic i/o
+// timeout, context deadline) are NOT retried: a completion POST is
+// non-idempotent and may already have reached and been processed by the server,
+// so replaying it could duplicate (billable) work. Only the INITIAL request is
+// ever in scope; once the response body starts streaming it is never re-issued.
+// Pre-send classification is by syscall errno (portable across POSIX and
+// Windows), scoped to an Op=="dial" *net.OpError so the same errno raised on a
+// post-send read/write can't be misread as pre-send. NXDOMAIN is treated as
+// permanent (not retried), and the pre-send path fails fast on its own
+// sub-second schedule rather than the seconds-long 429 backoff.
 
 const defaultMaxRetryAttempts = 6
 
@@ -37,12 +49,32 @@ const maxBackoff = 30 * time.Second
 // schedule is 2s, 4s, 8s, 16s, then maxBackoff. A var so tests can shrink it.
 var retryBackoffBase = 2 * time.Second
 
+// preSendMaxAttempts bounds retries of a provably pre-send transport failure.
+// Unlike a 429 (whose window is measured in seconds), a refused/unreachable dial
+// either recovers within a few hundred ms or is a deterministic misconfiguration
+// (wrong host/port, dead local daemon) that will never recover — so a short,
+// low-count schedule fails fast instead of stalling the agent ~60s on the
+// rate-limit schedule. This bounds the pre-send path independently of the caller's
+// maxAttempts, which is tuned for 429/503.
+const preSendMaxAttempts = 3
+
+// preSendBackoffBase is the first wait before replaying a pre-send transport
+// failure. Modeled on the mid-stream reconnect base (internal/agent/reconnect.go),
+// it is sub-second so a transient dial blip recovers quickly and a permanent one
+// fails fast. A var so tests can shrink it.
+var preSendBackoffBase = 500 * time.Millisecond
+
 // SendWithRetry issues an HTTP request, retrying ONLY the safe-to-replay server
 // responses (429 and 503, see ShouldRetryStatus) up to maxAttempts — backing off
 // between tries and honoring a server Retry-After header and context
-// cancellation. Other 5xx and transport/network errors are returned immediately,
-// never replayed (see the package note). The request is rebuilt from body each
-// attempt; setHeader (if non-nil) sets headers on every attempt.
+// cancellation. The one transport-error exception is a PROVABLY pre-send failure
+// (see isPreSendTransportError): a refused/unreachable dial or DNS failure where
+// no request bytes left this host, retried on its own short schedule
+// (preSendMaxAttempts / preSendBackoff). Every other 5xx and every ambiguous or
+// post-send transport error is returned immediately, never replayed, because a
+// non-idempotent completion POST may already have been received. The request is
+// rebuilt from body each attempt; setHeader (if non-nil) sets headers on every
+// attempt.
 //
 // It returns the final *http.Response (which the caller inspects for a non-2xx
 // status, exactly as before) or an error for a network failure / context
@@ -60,7 +92,47 @@ func SendWithRetry(
 	if maxAttempts <= 0 {
 		maxAttempts = defaultMaxRetryAttempts
 	}
-	for attempt := 1; ; attempt++ {
+	if client == nil {
+		client = http.DefaultClient
+	}
+	// Redirects are still FOLLOWED, so a redirecting gateway keeps working, but
+	// entering one makes that attempt non-retryable. The hazard is that client.Do
+	// can transmit the original POST to the first host, follow a 307/308, and then
+	// fail the dial to the redirect target: that error is an Op=="dial" error, so
+	// isPreSendTransportError would read it as "nothing was sent" and replay a
+	// completion the first host already received. Observing the redirect rather
+	// than blocking it keeps both properties, since a request that was redirected
+	// is never replayed while redirecting endpoints stay usable. The caller's own
+	// CheckRedirect still runs and still decides whether to proceed.
+	redirected := false
+	callerCheckRedirect := client.CheckRedirect
+	redirectAwareClient := *client
+	redirectAwareClient.CheckRedirect = func(request *http.Request, via []*http.Request) error {
+		redirected = true
+		if callerCheckRedirect != nil {
+			return callerCheckRedirect(request, via)
+		}
+		// Mirror net/http's default policy, which a nil CheckRedirect would have
+		// applied: stop after 10 hops.
+		if len(via) >= 10 {
+			return errors.New("stopped after 10 redirects")
+		}
+		return nil
+	}
+	// The two retry budgets are fully independent, each with its own counter and
+	// its own backoff ordinal, because they answer different questions: a
+	// dial/DNS/TLS failure is a fast local blip (preSendMaxAttempts, sub-second
+	// schedule) while a 429/503 is a server-side window measured in seconds
+	// (maxAttempts, retryBackoffBase). A single shared loop counter coupled them
+	// in both directions: status retries starved the pre-send budget, and
+	// pre-send retries shortened the status window and started its backoff at a
+	// later rung. Neither kind of failure now consumes or advances the other.
+	preSendAttempts := 0
+	statusAttempts := 0
+	for {
+		// Per attempt: only a redirect entered by THIS attempt can make its own
+		// transport error unsafe to replay.
+		redirected = false
 		request, err := http.NewRequestWithContext(ctx, method, url, bytes.NewReader(body))
 		if err != nil {
 			return nil, err
@@ -70,32 +142,78 @@ func SendWithRetry(
 		}
 
 		connectSpan := trace.FromContext(ctx).Span(trace.SpanProviderConnect)
-		response, err := client.Do(request)
+		response, err := redirectAwareClient.Do(request)
 		connectSpan.End()
 		if err != nil {
-			// A transport failure on a POST does NOT mean the server didn't receive
-			// it — the request may have arrived and be generating a (billable,
-			// non-idempotent) completion while only the response/connection failed.
-			// Replaying it could duplicate that work, so surface the error instead
-			// of retrying. Only 429/503 responses (below), where we KNOW the request
-			// was not accepted, are safe to retry.
+			// Context cancellation always surfaces as cancellation, never a retry.
 			if ctx.Err() != nil {
 				return nil, ctx.Err()
 			}
+			// A transport failure on a POST does NOT usually mean the server didn't
+			// receive it — the request may have arrived and be generating a
+			// (billable, non-idempotent) completion while only the response or
+			// connection failed. Replaying that could duplicate work, so it is
+			// surfaced immediately. The one safe exception is a PROVABLY pre-send
+			// failure (see isPreSendTransportError): the request never left this
+			// host, so a replay cannot duplicate anything — exactly like a 429/503
+			// where we KNOW the request was not accepted. The pre-send path uses
+			// its own short bound and sub-second schedule (preSendMaxAttempts /
+			// preSendBackoff), not the seconds-long 429 schedule, so a permanent
+			// dial failure fails fast instead of stalling the agent for ~60s.
+			// A redirect means the original POST already left this host, so the
+			// dial that failed belongs to a later hop and replaying would duplicate
+			// a completion the first host may already have processed.
+			if isPreSendTransportError(err) && !redirected {
+				preSendAttempts++
+				if preSendAttempts < preSendMaxAttempts {
+					if r := trace.FromContext(ctx); r != nil {
+						r.Counter(trace.CounterRetryCount, 1)
+					}
+					if preSendBackoff(ctx, preSendAttempts) {
+						continue
+					}
+					return nil, ctx.Err()
+				}
+			}
 			return nil, err
 		}
 
-		if ShouldRetryStatus(response.StatusCode) && attempt < maxAttempts {
-			if r := trace.FromContext(ctx); r != nil {
-				r.Counter(trace.CounterRetryCount, 1)
+		// Deliberately NOT guarded on `redirected`, unlike the transport-error
+		// branch above. The two differ in what is known about the request. A
+		// transport failure after a redirect leaves it unknown whether the
+		// redirect target received and began processing the POST, so replaying
+		// could duplicate billable work. Here every hop answered: a 307/308 is the
+		// origin declining and delegating, and a retryable status is the target
+		// declining as well, so nothing accepted the request and a replay
+		// duplicates nothing.
+		//
+		// The exposure is also identical to a retryable status with no redirect in
+		// play, which this package has always retried. Guarding on redirects here
+		// would not close a gap, it would make a rate-limited completion behind a
+		// redirecting gateway fail outright instead of backing off.
+		// The transport hop succeeded, so the connect phase is healthy again and
+		// the pre-send budget starts over. It bounds CONSECUTIVE connect failures,
+		// not the lifetime of the call: without this reset, two early dial retries
+		// followed by a 429 and then one more dial blip would surface a transport
+		// error while the status budget still had room, which is the coupling
+		// these two counters were split apart to remove. Total work stays bounded,
+		// since every reset costs a status attempt and those are capped.
+		preSendAttempts = 0
+
+		if ShouldRetryStatus(response.StatusCode) {
+			statusAttempts++
+			if statusAttempts < maxAttempts {
+				if r := trace.FromContext(ctx); r != nil {
+					r.Counter(trace.CounterRetryCount, 1)
+				}
+				wait := RetryAfter(response)
+				_ = response.Body.Close()
+				if Backoff(ctx, statusAttempts, wait) {
+					continue
+				}
+				// Backoff aborted: the only reason it returns false is ctx cancellation.
+				return nil, ctx.Err()
 			}
-			wait := RetryAfter(response)
-			_ = response.Body.Close()
-			if Backoff(ctx, attempt, wait) {
-				continue
-			}
-			// Backoff aborted: the only reason it returns false is ctx cancellation.
-			return nil, ctx.Err()
 		}
 
 		// Success, a non-retryable status, or retries exhausted on a retryable
@@ -109,6 +227,119 @@ func SendWithRetry(
 	}
 }
 
+// isPreSendTransportError reports whether a transport error PROVES no request
+// bytes reached the server, so replaying the request cannot duplicate a
+// billable completion. Only narrow, unambiguous pre-connect failures qualify: a
+// recoverable DNS lookup failure, a refused/unreachable dial, and a TLS
+// handshake timeout (the handshake completes before any HTTP request bytes are
+// written).
+//
+// The refused/unreachable errnos (ECONNREFUSED/ENETUNREACH/EHOSTUNREACH) are
+// matched by syscall errno via errors.Is — portable across platforms, since Go
+// maps the WSA* codes on Windows too — but ONLY when carried by an Op=="dial"
+// *net.OpError. errors.Is matches an errno anywhere in the chain regardless of
+// which socket operation raised it, and the kernel raises these same errnos on
+// an established connection when a route drops mid-generation; scoping to the
+// dial phase is what makes "no request bytes were sent" structural rather than
+// argued, so a post-send read/write can never replay a non-idempotent POST.
+// NXDOMAIN is treated as permanent (deterministic) and is NOT retried.
+//
+// Ambiguous or post-send failures are deliberately excluded and checked FIRST,
+// so an error that is both is never treated as pre-send: a reset or broken pipe
+// can follow a request that was already sent, and a bare "i/o timeout" covers
+// read timeouts after send as well as dial timeouts, so none of these prove the
+// request was not received and a non-idempotent POST must not be replayed on
+// them.
+func isPreSendTransportError(err error) bool {
+	if err == nil {
+		return false
+	}
+	// DNS resolution precedes any connection, so a lookup failure proves nothing
+	// was sent — but only retry a lookup that could plausibly succeed on a replay.
+	// NXDOMAIN (IsNotFound) is authoritative and deterministic (a mistyped or
+	// non-existent host), so retrying it just stalls the agent through the whole
+	// backoff schedule before failing anyway; a timeout or temporary SERVFAIL is
+	// worth another try. errors.As unwraps url.Error/OpError to reach the DNSError.
+	var dnsErr *net.DNSError
+	if errors.As(err, &dnsErr) {
+		return !dnsErr.IsNotFound
+	}
+	// Post-send / ambiguous failures, excluded first. EOF and reset are matched
+	// by identity so wording (or a hostname containing "eof") can't fool them.
+	if errors.Is(err, io.EOF) || errors.Is(err, io.ErrUnexpectedEOF) || isConnResetErrno(err) {
+		return false
+	}
+	// A CONNECT-phase timeout is provably pre-send for the same structural reason
+	// a refused dial is: the connection was never established, so no request
+	// bytes left this host. It has to be admitted BEFORE the wording exclusions
+	// below, because net surfaces it with "i/o timeout" in the message, which the
+	// blanket exclusion would otherwise reject. The gate is what keeps that safe:
+	// a post-send read timeout is also Timeout(), but carries Op "read", so it
+	// never reaches here.
+	if opErr := connectOpError(err); opErr != nil && opErr.Timeout() {
+		return true
+	}
+	msg := strings.ToLower(err.Error())
+	if strings.Contains(msg, "connection reset") ||
+		strings.Contains(msg, "broken pipe") ||
+		strings.Contains(msg, "i/o timeout") {
+		return false
+	}
+	// A refused/unreachable connection is provably pre-send ONLY when it was
+	// raised during the DIAL (connection-establishment) phase. The same errnos
+	// are also raised by the kernel on an ESTABLISHED connection — when a route
+	// drops or an ICMP unreachable arrives mid-generation — which is post-send.
+	// errors.Is matches an errno anywhere in the chain regardless of the
+	// operation, so we first require an Op=="dial" *net.OpError: that makes "no
+	// request bytes left this host" structural, so a post-send read/write carrying
+	// one of these can never replay a non-idempotent POST. dialPreSendErrnos is
+	// platform-specific: on Windows a refused dial carries the raw WSA* errno
+	// (WSAECONNREFUSED etc.), which does NOT satisfy errors.Is against the POSIX
+	// syscall.ECONNREFUSED, so the Windows list adds those codes. The string
+	// markers only catch a POSIX dial error already flattened past its errno.
+	if connectOpError(err) != nil {
+		for _, errno := range dialPreSendErrnos {
+			if errors.Is(err, errno) {
+				return true
+			}
+		}
+		switch {
+		case strings.Contains(msg, "connection refused"),
+			strings.Contains(msg, "network is unreachable"),
+			strings.Contains(msg, "no route to host"):
+			return true
+		}
+	}
+	// A TLS handshake times out after the TCP dial connects but before any HTTP
+	// request bytes are written, so it is structurally pre-send. net/http surfaces
+	// it as its own error string, not a *net.OpError, so it is matched by wording
+	// independent of the dial gate.
+	if strings.Contains(msg, "tls handshake timeout") {
+		return true
+	}
+	return false
+}
+
+// connectOpError returns the outermost *net.OpError in err's chain when it comes
+// from the connection-establishment phase, else nil. Go tags a direct dial with
+// Op "dial" and a failed CONNECT through an HTTP proxy with Op "proxyconnect";
+// a failure on an already-established connection carries Op "read"/"write".
+// This is how isPreSendTransportError scopes the pre-send match to the connect
+// phase so a post-send failure can't be misclassified as pre-send.
+//
+// "proxyconnect" belongs here for the same reason "dial" does: when HTTPS_PROXY
+// is set and the CONNECT fails, the tunnel to the model host was never built, so
+// no request bytes reached it. Omitting it meant every pre-send failure behind a
+// corporate proxy fell through to no-retry, which is the environment where
+// transient connect failures are most common.
+func connectOpError(err error) *net.OpError {
+	var opErr *net.OpError
+	if errors.As(err, &opErr) && (opErr.Op == "dial" || opErr.Op == "proxyconnect") {
+		return opErr
+	}
+	return nil
+}
+
 // ShouldRetryStatus reports whether an HTTP status is safe to retry for a
 // non-idempotent completion POST: 429 (Too Many Requests), 503 (Service
 // Unavailable), and 529 (Anthropic's "overloaded"). All mean the server
@@ -137,6 +368,39 @@ func Backoff(ctx context.Context, attempt int, retryAfter time.Duration) bool {
 	}
 }
 
+// preSendBackoff waits before pre-send retry attempt N (1-based) on a short
+// sub-second schedule (500ms, 1s, ...), returning false if the context is
+// cancelled during the wait. Separate from Backoff's seconds-long rate-limit
+// schedule because a pre-send dial failure recovers in milliseconds or never.
+func preSendBackoff(ctx context.Context, attempt int) bool {
+	timer := time.NewTimer(preSendBackoffWait(attempt))
+	defer timer.Stop()
+	select {
+	case <-ctx.Done():
+		return false
+	case <-timer.C:
+		return true
+	}
+}
+
+// preSendBackoffWait computes the pre-send wait before retry attempt N (1-based):
+// exponential from preSendBackoffBase (500ms, 1s, 2s, ...), capped at maxBackoff.
+// The exponent is clamped so a large attempt count cannot overflow.
+func preSendBackoffWait(attempt int) time.Duration {
+	exponent := attempt - 1
+	if exponent > 5 {
+		exponent = 5
+	}
+	if exponent < 0 {
+		exponent = 0
+	}
+	wait := preSendBackoffBase * time.Duration(1<<exponent)
+	if wait > maxBackoff {
+		wait = maxBackoff
+	}
+	return wait
+}
+
 // backoffWait computes the wait before retry attempt N (1-based): Retry-After
 // when supplied, else exponential from retryBackoffBase, both capped at
 // maxBackoff. The exponent is clamped so a large attempt count cannot overflow.

internal/providers/providerio/retry_dialerrno.go

@@ -0,0 +1,32 @@
+//go:build !windows && !plan9
+
+package providerio
+
+import (
+	"errors"
+	"syscall"
+)
+
+// dialPreSendErrnos are the syscall errnos a refused or unreachable DIAL raises
+// on this platform. On POSIX the standard syscall constants are exactly what the
+// net package surfaces, so matching them via errors.Is is sufficient. Plan 9 is
+// excluded because it does not define these BSD-style errno constants (it uses
+// string errors), so the constants below would not compile there; Zero targets
+// linux/darwin/windows, so no Plan 9 variant is needed.
+var dialPreSendErrnos = []error{
+	syscall.ECONNREFUSED,
+	syscall.ENETUNREACH,
+	syscall.EHOSTUNREACH,
+	// A connect timeout is normally caught by the Timeout() check on the connect
+	// gate; this covers the case where the errno survives but the timeout flag
+	// does not, e.g. an error rebuilt from the errno alone.
+	syscall.ETIMEDOUT,
+}
+
+// isConnResetErrno reports whether err carries a connection-reset errno, the
+// post-send exclusion in isPreSendTransportError. It lives here (per platform)
+// so retry.go references no syscall constants and stays buildable on Plan 9,
+// which defines none of them.
+func isConnResetErrno(err error) bool {
+	return errors.Is(err, syscall.ECONNRESET)
+}

internal/providers/providerio/retry_dialerrno_plan9.go

@@ -0,0 +1,15 @@
+//go:build plan9
+
+package providerio
+
+// Plan 9 has no BSD-style errno constants (it uses string errors), so the
+// errno-based classification is unavailable here: isConnResetErrno returns false
+// and dialPreSendErrnos is empty. retry.go still classifies via its string
+// markers on this platform, the pre-send inclusions (connection refused /
+// network unreachable / no route to host) and the separate post-send
+// "connection reset" exclusion each keep working through the wording checks.
+// Zero targets linux/darwin/windows, so this variant only keeps the package
+// building on Plan 9 rather than adding real support.
+var dialPreSendErrnos = []error{}
+
+func isConnResetErrno(error) bool { return false }

internal/providers/providerio/retry_dialerrno_windows.go

@@ -0,0 +1,40 @@
+//go:build windows
+
+package providerio
+
+import (
+	"errors"
+	"syscall"
+
+	"golang.org/x/sys/windows"
+)
+
+// dialPreSendErrnos are the errnos a refused or unreachable DIAL raises on
+// Windows. A real Windows dial failure carries the raw Winsock error
+// (WSAECONNREFUSED = 10061, WSAENETUNREACH = 10051, WSAEHOSTUNREACH = 10065), and
+// errors.Is against the POSIX syscall.ECONNREFUSED returns FALSE there because
+// Go's Windows syscall.ECONNREFUSED is a distinct value the net package never
+// produces. So the WSA codes must be matched explicitly, or a refused dial would
+// silently never be retried on Windows. The POSIX constants stay in the list too
+// so an error built with them (e.g. a test fixture) still classifies.
+var dialPreSendErrnos = []error{
+	syscall.ECONNREFUSED,
+	syscall.ENETUNREACH,
+	syscall.EHOSTUNREACH,
+	syscall.ETIMEDOUT,
+	windows.WSAECONNREFUSED,
+	windows.WSAENETUNREACH,
+	windows.WSAEHOSTUNREACH,
+	// Same reasoning as the POSIX ETIMEDOUT entry, with the Winsock code a real
+	// Windows connect timeout actually carries.
+	windows.WSAETIMEDOUT,
+}
+
+// isConnResetErrno reports whether err carries a connection-reset errno, the
+// post-send exclusion in isPreSendTransportError. It matches the same
+// syscall.ECONNRESET the shared path used before the errno classification was
+// split out per platform, so behavior on Windows is unchanged; it lives here
+// so retry.go references no syscall constants and stays buildable on Plan 9.
+func isConnResetErrno(err error) bool {
+	return errors.Is(err, syscall.ECONNRESET)
+}

internal/providers/providerio/retry_test.go

@@ -1,15 +1,34 @@
+//go:build !plan9
+
 package providerio
 
 import (
 	"context"
 	"errors"
+	"io"
+	"net"
 	"net/http"
 	"net/http/httptest"
+	"os"
 	"sync/atomic"
+	"syscall"
 	"testing"
 	"time"
 )
 
+// wrapDialErrno builds the error shape a dial failure has (a *net.OpError
+// wrapping an *os.SyscallError wrapping the errno) so errors.Is reaches the
+// errno exactly as in production. These fixtures use the POSIX syscall.Exxx
+// constants, which do NOT reproduce a real Windows dial: that carries distinct
+// windows.WSA* errnos (Go's syscall.ECONNREFUSED is a value the net package
+// never produces on Windows). They still classify on Windows only because
+// dialPreSendErrnos there also lists the POSIX constants for exactly this
+// fixture reason; the real Windows dial branch is covered separately by
+// TestIsPreSendTransportErrorRealRefusedDial, which makes an actual refused dial.
+func wrapDialErrno(op string, errno syscall.Errno) error {
+	return &net.OpError{Op: op, Net: "tcp", Err: os.NewSyscallError("connectex", errno)}
+}
+
 type roundTripperFunc func(*http.Request) (*http.Response, error)
 
 func (f roundTripperFunc) RoundTrip(r *http.Request) (*http.Response, error) { return f(r) }
@@ -35,6 +54,152 @@ func TestSendWithRetryDoesNotReplayTransportErrors(t *testing.T) {
 	}
 }
 
+// A PROVABLY pre-send transport failure (no request bytes left this host) is
+// safe to replay and must be retried, bounded by preSendMaxAttempts (its own
+// short schedule), unlike an ambiguous post-send failure. Real dial failures
+// arrive as an Op=="dial" *net.OpError, so the errno cases are the production
+// shape on every platform (Windows included); the string case exercises the
+// wording fallback for a dial error already flattened past its errno.
+func TestSendWithRetryReplaysProvablyPreSendErrors(t *testing.T) {
+	shrinkBackoff(t)
+	cases := map[string]error{
+		"errno refused (dial)":      wrapDialErrno("dial", syscall.ECONNREFUSED),
+		"errno network unreachable": wrapDialErrno("dial", syscall.ENETUNREACH),
+		"errno host unreachable":    wrapDialErrno("dial", syscall.EHOSTUNREACH),
+		"dial string fallback":      &net.OpError{Op: "dial", Net: "tcp", Err: errors.New("connect: connection refused")},
+		"tls handshake timeout":     errors.New("net/http: TLS handshake timeout"),
+		"dns timeout":               &net.DNSError{Err: "server misbehaving", Name: "nope.invalid", IsTimeout: true},
+	}
+	for name, transportErr := range cases {
+		t.Run(name, func(t *testing.T) {
+			var calls int32
+			client := &http.Client{Transport: roundTripperFunc(func(*http.Request) (*http.Response, error) {
+				atomic.AddInt32(&calls, 1)
+				return nil, transportErr
+			})}
+			// maxAttempts=6 (the default) proves the pre-send path is bounded by
+			// preSendMaxAttempts, not the caller's 429-tuned maxAttempts.
+			resp, err := SendWithRetry(context.Background(), client, http.MethodPost, "http://example.invalid", []byte("{}"), nil, 6)
+			if resp != nil {
+				_ = resp.Body.Close()
+			}
+			if err == nil {
+				t.Fatal("expected the transport error to surface after retries are exhausted")
+			}
+			if got := atomic.LoadInt32(&calls); got != preSendMaxAttempts {
+				t.Fatalf("pre-send error tried %d times, want %d (preSendMaxAttempts)", got, preSendMaxAttempts)
+			}
+		})
+	}
+}
+
+// An ambiguous transport failure that could have followed a sent request must
+// NOT be replayed: a non-idempotent completion POST could duplicate billable
+// work. This is the safety line the fix must not cross.
+func TestSendWithRetryDoesNotReplayAmbiguousTransportErrors(t *testing.T) {
+	for name, transportErr := range map[string]error{
+		"generic i/o timeout": errors.New("dial tcp 1.2.3.4:443: i/o timeout"),
+		"broken pipe":         errors.New("write tcp: broken pipe"),
+		"unexpected eof":      io.ErrUnexpectedEOF,
+		// The pre-send errnos are also raised on an ESTABLISHED connection: a route
+		// dropping mid-generation surfaces on the pending read/write, which is
+		// post-send. Scoping to Op=="dial" must keep these from replaying the POST.
+		"host unreachable on read is post-send": wrapDialErrno("read", syscall.EHOSTUNREACH),
+		"net unreachable on write is post-send": wrapDialErrno("write", syscall.ENETUNREACH),
+		// NXDOMAIN is authoritative and deterministic, so retrying it would only
+		// stall the agent before failing anyway.
+		"dns nxdomain": &net.DNSError{Err: "no such host", Name: "nope.invalid", IsNotFound: true},
+	} {
+		t.Run(name, func(t *testing.T) {
+			var calls int32
+			client := &http.Client{Transport: roundTripperFunc(func(*http.Request) (*http.Response, error) {
+				atomic.AddInt32(&calls, 1)
+				return nil, transportErr
+			})}
+			resp, err := SendWithRetry(context.Background(), client, http.MethodPost, "http://example.invalid", []byte("{}"), nil, 3)
+			if resp != nil {
+				_ = resp.Body.Close()
+			}
+			if err == nil {
+				t.Fatal("expected the transport error to surface")
+			}
+			if got := atomic.LoadInt32(&calls); got != 1 {
+				t.Fatalf("ambiguous transport error replayed %d times, want 1 (no retry)", got)
+			}
+		})
+	}
+}
+
+// A REAL refused dial (nothing listening on the target) must classify as
+// pre-send on every platform. This is the regression the errno-constant fixtures
+// miss: on Windows the kernel raises WSAECONNREFUSED, which errors.Is does NOT
+// match against syscall.ECONNREFUSED, so before dialPreSendErrnos carried the WSA
+// codes a refused dial returned false here and Windows never retried it. Using a
+// real Dial exercises the platform's true error shape.
+func TestIsPreSendTransportErrorRealRefusedDial(t *testing.T) {
+	// Reserve an ephemeral port, then close it, so a dial there is refused.
+	ln, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatalf("listen: %v", err)
+	}
+	addr := ln.Addr().String()
+	if cerr := ln.Close(); cerr != nil {
+		t.Fatalf("close listener: %v", cerr)
+	}
+	conn, dialErr := (&net.Dialer{Timeout: 2 * time.Second}).Dial("tcp", addr)
+	if dialErr == nil {
+		_ = conn.Close()
+		t.Skip("expected a refused dial but the port accepted a connection")
+	}
+	if !isPreSendTransportError(dialErr) {
+		t.Fatalf("real refused dial not classified pre-send: %v", dialErr)
+	}
+}
+
+func TestIsPreSendTransportError(t *testing.T) {
+	cases := []struct {
+		name string
+		err  error
+		want bool
+	}{
+		{"nil", nil, false},
+		// DNS: a lookup that could recover is pre-send; NXDOMAIN is permanent.
+		{"dns timeout", &net.DNSError{Err: "server misbehaving", Name: "x.invalid", IsTimeout: true}, true},
+		{"dns temporary", &net.DNSError{Err: "server misbehaving", Name: "x.invalid", IsTemporary: true}, true},
+		{"dns nxdomain is permanent", &net.DNSError{Err: "no such host", Name: "x.invalid", IsNotFound: true}, false},
+		{"tls handshake timeout", errors.New("net/http: TLS handshake timeout"), true},
+		// Errno-wrapped DIAL failures: how a real refused/unreachable dial arrives
+		// on EVERY platform, including Windows where the wording differs entirely.
+		{"errno refused (dial)", wrapDialErrno("dial", syscall.ECONNREFUSED), true},
+		{"errno network unreachable (dial)", wrapDialErrno("dial", syscall.ENETUNREACH), true},
+		{"errno host unreachable (dial)", wrapDialErrno("dial", syscall.EHOSTUNREACH), true},
+		{"dial operror string fallback", &net.OpError{Op: "dial", Net: "tcp", Err: errors.New("connect: connection refused")}, true},
+		// The SAME errnos raised on a post-send read/write are NOT pre-send — the
+		// kernel reports them on an established connection when a route drops.
+		{"host unreachable on read is post-send", wrapDialErrno("read", syscall.EHOSTUNREACH), false},
+		{"net unreachable on write is post-send", wrapDialErrno("write", syscall.ENETUNREACH), false},
+		{"errno reset is post-send", wrapDialErrno("read", syscall.ECONNRESET), false},
+		// A refused/unreachable error already flattened past its *net.OpError can't
+		// be proven pre-send, so it is NOT retried (conservative direction).
+		{"flattened refused, no opError", errors.New("dial tcp 127.0.0.1:1: connect: connection refused"), false},
+		{"connection reset", errors.New("read tcp: connection reset by peer"), false},
+		{"broken pipe", errors.New("write tcp: broken pipe"), false},
+		{"unexpected eof", io.ErrUnexpectedEOF, false},
+		{"eof", io.EOF, false},
+		{"generic io timeout", errors.New("dial tcp: i/o timeout"), false},
+		{"context deadline", context.DeadlineExceeded, false},
+		// Exclusion is checked before inclusion, even for a dial OpError: an "i/o
+		// timeout" in the message wins over the refused wording.
+		{"exclusion wins over inclusion", &net.OpError{Op: "dial", Net: "tcp", Err: errors.New("connection refused / i/o timeout")}, false},
+		{"unrelated", errors.New("some other error"), false},
+	}
+	for _, c := range cases {
+		if got := isPreSendTransportError(c.err); got != c.want {
+			t.Errorf("%s: isPreSendTransportError(%v) = %v, want %v", c.name, c.err, got, c.want)
+		}
+	}
+}
+
 func TestShouldRetryStatus(t *testing.T) {
 	cases := map[int]bool{
 		http.StatusOK:                  false,
@@ -94,12 +259,17 @@ func TestBackoffWaitsThenReturnsTrue(t *testing.T) {
 	}
 }
 
-// shrinkBackoff makes retry waits negligible for the duration of a test.
+// shrinkBackoff makes both retry schedules (429 and pre-send) negligible for the
+// duration of a test.
 func shrinkBackoff(t *testing.T) {
 	t.Helper()
-	saved := retryBackoffBase
+	savedRetry, savedPreSend := retryBackoffBase, preSendBackoffBase
 	retryBackoffBase = time.Millisecond
-	t.Cleanup(func() { retryBackoffBase = saved })
+	preSendBackoffBase = time.Millisecond
+	t.Cleanup(func() {
+		retryBackoffBase = savedRetry
+		preSendBackoffBase = savedPreSend
+	})
 }
 
 func TestBackoffWaitSchedule(t *testing.T) {
@@ -126,6 +296,30 @@ func TestBackoffWaitSchedule(t *testing.T) {
 	}
 }
 
+// The pre-send schedule is sub-second and doubles, far shorter than the 429
+// schedule: a permanent dial failure fails in ~1.5s across preSendMaxAttempts
+// (500ms + 1s) instead of stalling the agent ~60s on 2/4/8/16/30s. This is the
+// second half of the fix for a mistyped host / dead local daemon hanging a turn.
+func TestPreSendBackoffWaitSchedule(t *testing.T) {
+	cases := []struct {
+		attempt int
+		want    time.Duration
+	}{
+		{1, 500 * time.Millisecond},
+		{2, 1 * time.Second},
+		{3, 2 * time.Second},
+		// The exponent clamps at 5 (500ms<<5 = 16s) so a large attempt count can't
+		// overflow; in practice preSendMaxAttempts caps retries at attempt 2, so
+		// only the 500ms/1s rungs are ever reached.
+		{50, 16 * time.Second},
+	}
+	for _, c := range cases {
+		if got := preSendBackoffWait(c.attempt); got != c.want {
+			t.Errorf("preSendBackoffWait(%d) = %v, want %v", c.attempt, got, c.want)
+		}
+	}
+}
+
 func TestSendWithRetryRetriesThenSucceeds(t *testing.T) {
 	shrinkBackoff(t)
 	var hits int32
@@ -193,3 +387,316 @@ func TestSendWithRetryReturnsLastResponseAfterMaxAttempts(t *testing.T) {
 		t.Fatalf("server hit %d times, want 2 (maxAttempts)", got)
 	}
 }
+
+// Redirects are still followed, so a gateway that answers the completion endpoint
+// with a 307/308 keeps working: the caller gets the final response, not the 3xx.
+func TestSendWithRetryFollowsRedirects(t *testing.T) {
+	var calls int32
+	client := &http.Client{Transport: roundTripperFunc(func(r *http.Request) (*http.Response, error) {
+		if atomic.AddInt32(&calls, 1) == 1 {
+			return &http.Response{
+				StatusCode: http.StatusTemporaryRedirect,
+				Header:     http.Header{"Location": {"https://redirect-target.invalid/v1"}},
+				Body:       http.NoBody,
+				Request:    r,
+			}, nil
+		}
+		return &http.Response{StatusCode: http.StatusOK, Body: http.NoBody, Request: r}, nil
+	})}
+
+	resp, err := SendWithRetry(context.Background(), client, http.MethodPost, "https://origin.invalid/v1", []byte("{}"), nil, 3)
+	if err != nil {
+		t.Fatalf("a redirected completion must succeed: %v", err)
+	}
+	_ = resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("status = %d, want 200 (the redirect should have been followed)", resp.StatusCode)
+	}
+	if got := atomic.LoadInt32(&calls); got != 2 {
+		t.Fatalf("transport called %d times, want 2 (origin then redirect target)", got)
+	}
+}
+
+// Once an attempt has entered a redirect, the original POST has already left this
+// host, so a dial failure on the redirect hop must NOT be replayed even though it
+// arrives as an Op=="dial" error that isPreSendTransportError would otherwise
+// treat as provably pre-send. Replaying it would re-bill a completion the first
+// host may already have processed.
+func TestSendWithRetryDoesNotReplayAfterRedirect(t *testing.T) {
+	shrinkBackoff(t)
+	var calls int32
+	client := &http.Client{Transport: roundTripperFunc(func(r *http.Request) (*http.Response, error) {
+		if atomic.AddInt32(&calls, 1) == 1 {
+			return &http.Response{
+				StatusCode: http.StatusTemporaryRedirect,
+				Header:     http.Header{"Location": {"https://redirect-target.invalid/v1"}},
+				Body:       http.NoBody,
+				Request:    r,
+			}, nil
+		}
+		// The dial to the redirect target fails: pre-send in shape, but post-send
+		// in fact, because the origin already received the POST.
+		return nil, wrapDialErrno("dial", syscall.ECONNREFUSED)
+	})}
+
+	resp, err := SendWithRetry(context.Background(), client, http.MethodPost, "https://origin.invalid/v1", []byte("{}"), nil, 6)
+	if resp != nil {
+		_ = resp.Body.Close()
+	}
+	if err == nil {
+		t.Fatal("expected the redirect-hop dial failure to surface as an error")
+	}
+	// Exactly two: the origin request and the single redirect hop. A third would
+	// mean the redirected POST was replayed.
+	if got := atomic.LoadInt32(&calls); got != 2 {
+		t.Fatalf("redirected POST was replayed: transport called %d times, want 2", got)
+	}
+}
+
+// A caller that supplies its own CheckRedirect keeps it: the retry path observes
+// redirects, it does not take the decision away from the caller.
+func TestSendWithRetryPreservesCallerRedirectPolicy(t *testing.T) {
+	var policyCalls int32
+	client := &http.Client{
+		CheckRedirect: func(*http.Request, []*http.Request) error {
+			atomic.AddInt32(&policyCalls, 1)
+			return http.ErrUseLastResponse
+		},
+		Transport: roundTripperFunc(func(r *http.Request) (*http.Response, error) {
+			return &http.Response{
+				StatusCode: http.StatusTemporaryRedirect,
+				Header:     http.Header{"Location": {"https://redirect-target.invalid/v1"}},
+				Body:       http.NoBody,
+				Request:    r,
+			}, nil
+		}),
+	}
+
+	resp, err := SendWithRetry(context.Background(), client, http.MethodPost, "https://origin.invalid/v1", []byte("{}"), nil, 3)
+	if err != nil {
+		t.Fatalf("caller policy asked to stop at the 3xx, so it must surface: %v", err)
+	}
+	_ = resp.Body.Close()
+	if resp.StatusCode != http.StatusTemporaryRedirect {
+		t.Fatalf("status = %d, want the 307 surfaced per the caller's policy", resp.StatusCode)
+	}
+	if got := atomic.LoadInt32(&policyCalls); got != 1 {
+		t.Fatalf("caller CheckRedirect called %d times, want 1 (it must not be discarded)", got)
+	}
+}
+
+// The pre-send retry budget is independent of the 429/503 status retries: two
+// rate-limit responses must not consume the pre-send allowance, so the first
+// refused dial after them still gets its own preSendMaxAttempts tries.
+func TestSendWithRetryPreSendBudgetSurvivesStatusRetries(t *testing.T) {
+	shrinkBackoff(t)
+	var calls int32
+	client := &http.Client{Transport: roundTripperFunc(func(r *http.Request) (*http.Response, error) {
+		if atomic.AddInt32(&calls, 1) <= 2 {
+			return &http.Response{StatusCode: http.StatusTooManyRequests, Body: http.NoBody, Request: r}, nil
+		}
+		return nil, wrapDialErrno("dial", syscall.ECONNREFUSED)
+	})}
+
+	resp, err := SendWithRetry(context.Background(), client, http.MethodPost, "https://x.invalid/v1", []byte("{}"), nil, 6)
+	if resp != nil {
+		_ = resp.Body.Close()
+	}
+	if err == nil {
+		t.Fatal("expected the exhausted pre-send failure to surface as an error")
+	}
+	// Two status responses (calls 1-2), then the pre-send failure gets its own
+	// budget: preSendMaxAttempts=3 means it retries twice more before returning.
+	if got := atomic.LoadInt32(&calls); got != 2+preSendMaxAttempts {
+		t.Fatalf("pre-send budget eaten by status retries: transport called %d times, want %d (2 status + %d pre-send)", got, 2+preSendMaxAttempts, preSendMaxAttempts)
+	}
+}
+
+// The inverse sequence: pre-send retries must not consume the 429/503 budget or
+// advance its backoff rung either. Two safely replayed refused dials followed by
+// rate limiting must still leave the status path its full maxAttempts window,
+// otherwise a couple of dial blips silently shorten every later rate-limit wait.
+func TestSendWithRetryStatusBudgetSurvivesPreSendRetries(t *testing.T) {
+	shrinkBackoff(t)
+	const maxAttempts = 3
+	var calls int32
+	client := &http.Client{Transport: roundTripperFunc(func(r *http.Request) (*http.Response, error) {
+		if atomic.AddInt32(&calls, 1) <= 2 {
+			return nil, wrapDialErrno("dial", syscall.ECONNREFUSED)
+		}
+		return &http.Response{StatusCode: http.StatusTooManyRequests, Body: http.NoBody, Request: r}, nil
+	})}
+
+	resp, err := SendWithRetry(context.Background(), client, http.MethodPost, "https://x.invalid/v1", []byte("{}"), nil, maxAttempts)
+	if err != nil {
+		t.Fatalf("exhausted status retries surface the response, not an error: %v", err)
+	}
+	_ = resp.Body.Close()
+	if resp.StatusCode != http.StatusTooManyRequests {
+		t.Fatalf("status = %d, want 429", resp.StatusCode)
+	}
+	// Calls 1-2 are the retried dial failures; the status path then gets its own
+	// full window (maxAttempts requests) before the 429 is surfaced.
+	if got := atomic.LoadInt32(&calls); got != 2+maxAttempts {
+		t.Fatalf("status budget eaten by pre-send retries: transport called %d times, want %d (2 pre-send + %d status)", got, 2+maxAttempts, maxAttempts)
+	}
+}
+
+// A redirect followed by a RETRYABLE STATUS is still retried, and deliberately
+// so. This is the case a reviewer flagged as a possible replay of billable work,
+// and the asymmetry with the transport-error guard above is intentional rather
+// than an oversight, so it is pinned here.
+//
+// The two situations differ in what is known about the request:
+//
+//   - Transport error after a redirect: the original POST reached the origin and
+//     the dial to the redirect target then failed, so whether the target received
+//     and began processing it is UNKNOWN. Replaying could duplicate work, which
+//     is why that path refuses to retry.
+//   - Retryable status after a redirect: every hop answered. A 307/308 is the
+//     origin declining to process and delegating, and a 429/503/529 is the target
+//     declining as well. Nothing accepted the request, so replaying duplicates
+//     nothing.
+//
+// The risk here is also exactly the risk of a 429 without any redirect, which
+// this package has always retried; the redirect adds no new ambiguity. Refusing
+// to retry would instead make a rate-limited completion behind a redirecting
+// gateway fail outright rather than back off.
+func TestSendWithRetryRetriesRedirectedRetryableStatus(t *testing.T) {
+	var calls int32
+	client := &http.Client{Transport: roundTripperFunc(func(r *http.Request) (*http.Response, error) {
+		switch atomic.AddInt32(&calls, 1) {
+		case 1, 3:
+			// The origin declines and delegates, both times.
+			return &http.Response{
+				StatusCode: http.StatusTemporaryRedirect,
+				Header:     http.Header{"Location": {"https://redirect-target.invalid/v1"}},
+				Body:       http.NoBody,
+				Request:    r,
+			}, nil
+		case 2:
+			// The target declines too: rate limited, so nothing was accepted.
+			return &http.Response{StatusCode: http.StatusTooManyRequests, Body: http.NoBody, Request: r}, nil
+		default:
+			return &http.Response{StatusCode: http.StatusOK, Body: http.NoBody, Request: r}, nil
+		}
+	})}
+
+	resp, err := SendWithRetry(context.Background(), client, http.MethodPost, "https://origin.invalid/v1", []byte("{}"), nil, 3)
+	if err != nil {
+		t.Fatalf("a rate-limited completion behind a redirect must still retry: %v", err)
+	}
+	_ = resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("status = %d, want 200 after the retry succeeded", resp.StatusCode)
+	}
+	// origin, target(429), origin, target(200).
+	if got := atomic.LoadInt32(&calls); got != 4 {
+		t.Fatalf("transport called %d times, want 4 (redirect, 429, redirect, 200)", got)
+	}
+}
+
+// A CONNECT-phase timeout is provably pre-send: the connection was never
+// established, so no request bytes left this host. It reaches the classifier
+// carrying "i/o timeout", which the blanket wording exclusion rejects, so
+// admitting it depends on the connect gate running first.
+//
+// The post-send counterpart is the point of the table: a read timeout is also
+// Timeout(), and must stay excluded, or a reply that timed out mid-generation
+// would replay a completion the server already billed.
+func TestPreSendClassifiesConnectTimeouts(t *testing.T) {
+	timeoutErr := &timeoutError{}
+	for name, testCase := range map[string]struct {
+		err  error
+		want bool
+	}{
+		"dial timeout": {
+			err:  &net.OpError{Op: "dial", Net: "tcp", Err: timeoutErr},
+			want: true,
+		},
+		"proxy connect timeout": {
+			err:  &net.OpError{Op: "proxyconnect", Net: "tcp", Err: timeoutErr},
+			want: true,
+		},
+		"read timeout stays post-send": {
+			err:  &net.OpError{Op: "read", Net: "tcp", Err: timeoutErr},
+			want: false,
+		},
+		"write timeout stays post-send": {
+			err:  &net.OpError{Op: "write", Net: "tcp", Err: timeoutErr},
+			want: false,
+		},
+	} {
+		t.Run(name, func(t *testing.T) {
+			if got := isPreSendTransportError(testCase.err); got != testCase.want {
+				t.Fatalf("isPreSendTransportError(%v) = %v, want %v", testCase.err, got, testCase.want)
+			}
+		})
+	}
+}
+
+// A failed CONNECT through an HTTP proxy is pre-send for the same structural
+// reason a refused dial is, but net tags it Op "proxyconnect", so a gate
+// matching only "dial" left every pre-send failure behind a corporate proxy
+// unretried.
+func TestPreSendClassifiesProxyConnectRefusal(t *testing.T) {
+	err := &net.OpError{Op: "proxyconnect", Net: "tcp", Err: syscall.ECONNREFUSED}
+	if !isPreSendTransportError(err) {
+		t.Fatal("a refused proxy CONNECT was not classified pre-send; no request bytes reached the model host")
+	}
+	// The op still has to be a connect phase. The same errno on an established
+	// connection is post-send and must not replay a POST.
+	post := &net.OpError{Op: "read", Net: "tcp", Err: syscall.ECONNREFUSED}
+	if isPreSendTransportError(post) {
+		t.Fatal("a read-phase error was classified pre-send")
+	}
+}
+
+// timeoutError is a net.Error whose Timeout() is true, matching the shape
+// net/http surfaces for a connect deadline.
+type timeoutError struct{}
+
+func (timeoutError) Error() string   { return "i/o timeout" }
+func (timeoutError) Timeout() bool   { return true }
+func (timeoutError) Temporary() bool { return true }
+
+// The pre-send budget bounds CONSECUTIVE connect failures, not the lifetime of
+// the call. Without a reset on a successful hop, an early run of dial blips
+// permanently spends the budget, so a later blip surfaces as a transport error
+// while the status budget still has room. That is the coupling the two counters
+// were split apart to remove, in the one ordering the earlier tests did not
+// cover: pre-send, then status, then pre-send again.
+func TestPreSendBudgetResetsAfterSuccessfulHop(t *testing.T) {
+	refused := &net.OpError{Op: "dial", Net: "tcp", Err: syscall.ECONNREFUSED}
+	var calls int32
+	client := &http.Client{Transport: roundTripperFunc(func(r *http.Request) (*http.Response, error) {
+		switch atomic.AddInt32(&calls, 1) {
+		case 1, 2:
+			// Two connect failures, spending the pre-send budget under the old
+			// lifetime accounting.
+			return nil, refused
+		case 3:
+			// A hop that reaches the server. It declines, so the request was not
+			// accepted and retrying stays safe.
+			return &http.Response{StatusCode: http.StatusTooManyRequests, Body: http.NoBody, Request: r}, nil
+		case 4:
+			// One more connect blip. With a lifetime cap this surfaces as an
+			// error; the call should still recover.
+			return nil, refused
+		default:
+			return &http.Response{StatusCode: http.StatusOK, Body: http.NoBody, Request: r}, nil
+		}
+	})}
+
+	resp, err := SendWithRetry(context.Background(), client, http.MethodPost, "https://origin.invalid/v1", []byte("{}"), nil, 3)
+	if err != nil {
+		t.Fatalf("interleaved connect and status failures returned %v; the pre-send budget did not reset after the hop that reached the server", err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("status = %d, want 200", resp.StatusCode)
+	}
+	if got := atomic.LoadInt32(&calls); got != 5 {
+		t.Fatalf("transport called %d times, want 5 (dial, dial, 429, dial, 200)", got)
+	}
+}