test(tools): restore the drain guard and gate the TTY loop on the exit too

Two follow-ups from review.

The anti-redelivery poll my previous commit message described was not in the
tree. It was written, it passed, and then a `git checkout -- .` in the
falsification run that followed reverted it, because it was the one edit made
after the checkpoint commit. The message was amended afterwards and recorded work
that no longer existed. jatmn caught the mismatch. Restored here, and the claim
now matches the code: one more read after the served line, asserting it is not
repeated, which is the only guard for drain semantics on Windows since the
manager-level test that counts deliveries is skipped there.

The TTY loop stopped on the echo alone while asserting the exit as well. collect
can return the drained echo on a yield timer before markDone has run, so a poll
can carry got:hello with no exit_code meta yet and the assertion below would then
be about an exit nobody observed. It now waits for both. The echo term stays in
the condition deliberately: gating on exit_code alone would stop polling before a
slow shell start delivers the echo, which is the race this change exists to
remove.

internal/tools/bash_tool_test.go

@@ -70,6 +70,9 @@ func runBashToolHelper(command string) {
 		fmt.Println("listening", listener.Addr().String())
 		server := &http.Server{
 			Handler: http.HandlerFunc(func(response http.ResponseWriter, request *http.Request) {
+				// Printed per request, so a test can prove the session delivers output
+				// mid-flight: this line cannot exist before the first read has happened.
+				fmt.Println("served", request.URL.Path)
 				_, _ = response.Write([]byte("zero-server-ok"))
 			}),
 		}
@@ -522,7 +525,14 @@ func TestBashToolTimeoutKillsBackgroundChildren(t *testing.T) {
 	if result.Status != StatusError {
 		t.Fatalf("expected timeout error status, got %s: %q", result.Status, result.Output)
 	}
-	if elapsed > time.Second {
+	// The regression this guards is Run blocking until the background child's own
+	// sleep ends, so the bound only has to sit under that. One second sat almost
+	// exactly on it: the call legitimately costs a sandbox plan, a fork and exec of
+	// a shell plus a subshell, the 300ms timeout itself and the post-kill drain,
+	// and spawn alone was measured here at 720ms to 1.44s with every core busy. A
+	// merely slow runner therefore failed a correct system. Three seconds is still
+	// far below the child's lifetime, so nothing is given up.
+	if elapsed > 3*time.Second {
 		t.Fatalf("Run blocked %s past the 300ms timeout; background child held the pipes", elapsed)
 	}
 

internal/tools/exec_command_test.go

@@ -3,6 +3,7 @@ package tools
 import (
 	"context"
 	"io"
+	"net"
 	"net/http"
 	"os"
 	"path/filepath"
@@ -373,10 +374,6 @@ func TestExecCommandForegroundServerReturnsSessionAndServesHTTP(t *testing.T) {
 	if err != nil {
 		t.Fatalf("foreground server should return session_id, meta=%#v output=%q", start.Meta, start.Output)
 	}
-	addr := parseListeningAddress(start.Output)
-	if addr == "" {
-		t.Fatalf("server output did not include listening address: %q", start.Output)
-	}
 	t.Cleanup(func() {
 		writeTool.Run(context.Background(), map[string]any{
 			"session_id": sessionID,
@@ -384,7 +381,12 @@ func TestExecCommandForegroundServerReturnsSessionAndServesHTTP(t *testing.T) {
 		})
 	})
 
-	response, err := http.Get("http://" + addr)
+	addr := waitForListeningAddress(t, writeTool, sessionID, start.Output)
+
+	// Bounded, so a server that accepts a connection and never answers fails here
+	// with its address instead of hanging until the package timeout kills the run.
+	client := &http.Client{Timeout: 30 * time.Second}
+	response, err := client.Get("http://" + addr)
 	if err != nil {
 		t.Fatalf("foreground exec server was not reachable at %s: %v; output=%q", addr, err, start.Output)
 	}
@@ -396,6 +398,153 @@ func TestExecCommandForegroundServerReturnsSessionAndServesHTTP(t *testing.T) {
 	if string(bytes) != "zero-server-ok" {
 		t.Fatalf("server response = %q", string(bytes))
 	}
+
+	// AND THE SESSION DELIVERS OUTPUT MID-FLIGHT, PINNED DETERMINISTICALLY.
+	//
+	// The helper prints this line only when it serves a request, so it cannot have
+	// been in the first read: the request had not happened yet. It can only arrive
+	// through a poll of a process that is still running, which is the delivery the
+	// address parse used to cover by accident and only on a machine slow enough to
+	// miss the first read. A regression where a live session buffers everything
+	// until it exits leaves the rest of this test green and fails here.
+	deadline := time.Now().Add(60 * time.Second)
+	var served strings.Builder
+	for !strings.Contains(served.String(), "served") && time.Now().Before(deadline) {
+		poll := writeTool.Run(context.Background(), map[string]any{
+			"session_id":    sessionID,
+			"chars":         "",
+			"yield_time_ms": 250,
+		})
+		served.WriteString(poll.Output)
+		if poll.Status != StatusOK {
+			t.Fatalf("the exec session was gone before it reported serving the request:\n%s", served.String())
+		}
+	}
+	if !strings.Contains(served.String(), "served") {
+		t.Fatalf("a still-running session never delivered the line the server printed while serving:\n%s", served.String())
+	}
+
+	// AND A READ DRAINS WHAT IT RETURNS.
+	//
+	// One more poll, because "the line showed up" is satisfied just as well by
+	// output that is re-delivered forever. drain copies and then nils the buffer,
+	// so this read carries nothing; if it stopped clearing, every session in
+	// production would repeat its whole output on every poll while every other
+	// assertion here stayed green. The manager-level test that asserts the same
+	// count is skipped on Windows, so this is the only place guarding it there.
+	again := writeTool.Run(context.Background(), map[string]any{
+		"session_id":    sessionID,
+		"chars":         "",
+		"yield_time_ms": 250,
+	})
+	if strings.Contains(again.Output, "served") {
+		t.Errorf("a second read returned output the first had already delivered, so reads are no longer draining:\n%s", again.Output)
+	}
+}
+
+// waitForListeningAddress polls the exec session until the helper says where it
+// is listening.
+//
+// ONE YIELD IS AN ALLOWANCE, NOT A GUARANTEE, AND THIS TEST TREATED IT AS ONE.
+//
+// The helper is this test binary re-executed. The first read used to be its only
+// chance: 500ms to be spawned through the process manager, reach net.Listen and
+// have its line drained back. A Windows CI runner under load does not get there
+// in time, so the read returned "Command is still running." with no address and
+// the test failed as though the server were broken. It failed that way on
+// unrelated contributor PRs, which is worse than a slow test: it reports somebody
+// else's change as the fault.
+//
+// Polling is what the banner in that very output tells the caller to do, so it is
+// what this does, up to a bound far above any plausible start.
+//
+// AND IT STILL FAILS FAST WHEN THE FAILURE IS REAL. A session that has gone away
+// answers with StatusError, and a process that exited before listening answers
+// with an exit_code line. Either is a genuine defect, so neither is worth waiting
+// out: they end the loop immediately with what the session actually said.
+//
+// Each read DRAINS: managedProcess.collect returns only the bytes produced since
+// the previous call, so these chunks are incremental and none of them is the whole
+// output. Each is parsed on its own, which is sound because the helper writes its
+// address with a single Println and a write that small does not tear. The whole
+// transcript is kept anyway, so a failure shows every chunk rather than the last.
+func waitForListeningAddress(t *testing.T, writeTool Tool, sessionID int, first string) string {
+	t.Helper()
+	if addr := validListeningAddress(first); addr != "" {
+		return addr
+	}
+	// AND WHEN IT DOES NOT, THIS TEST STOPS COVERING THE FIRST READ.
+	//
+	// That parse used to be the only assertion in the repo that a still-running
+	// session's first read carries the child's stdout at all: every other test
+	// either uses a tiny yield against a helper that prints later, or uses a large
+	// one against a helper that exits, which is the collect-on-exit branch. The
+	// manager-level equivalent is skipped on Windows because it shells out to
+	// /bin/sh. The coverage cannot be pinned deterministically without
+	// reintroducing the race, since collect() blocks the whole window either way,
+	// so it is given up deliberately here and paid back after the request below,
+	// where a line that cannot exist before the first read proves the poll path.
+	transcript := []string{first}
+	// Sixty seconds, not thirty. Starting this helper was measured here at 0.74s
+	// to 4.79s with every core saturated, on a box faster than a CI runner, and
+	// windows-latest additionally scans a freshly linked test binary on its first
+	// execution. A deadline of the same order as the latency it tolerates is the
+	// mistake that produced the 500ms budget; the point is a bound, and every
+	// genuine failure below leaves before reaching it anyway.
+	deadline := time.Now().Add(60 * time.Second)
+	for time.Now().Before(deadline) {
+		poll := writeTool.Run(context.Background(), map[string]any{
+			"session_id":    sessionID,
+			"chars":         "",
+			"yield_time_ms": 250,
+		})
+		transcript = append(transcript, poll.Output)
+		if addr := validListeningAddress(poll.Output); addr != "" {
+			return addr
+		}
+		// FAST-FAIL, BECAUSE THE LOOP HAS NO PACING OF ITS OWN. All the spacing
+		// comes from collect() waiting out its yield inside a LIVE session; every
+		// error return happens in microseconds. Continue removes the process the
+		// moment it observes the exit, so without these two the loop free-runs on
+		// a dead id: measured at 3028 polls in 5 seconds, which on a 4-vCPU runner
+		// would peg a core for the whole deadline and hand t.Fatalf hundreds of
+		// kilobytes of the same sentence. The exit result is also delivered
+		// exactly once, so folding it into the blob and continuing loses the one
+		// poll that says what went wrong.
+		if poll.Status != StatusOK {
+			t.Fatalf("the exec session was gone before the server reported an address:\n%s", strings.Join(transcript, "\n--- poll ---\n"))
+		}
+		if strings.Contains(poll.Output, "exit_code:") {
+			t.Fatalf("the server process exited before it reported an address:\n%s", strings.Join(transcript, "\n--- poll ---\n"))
+		}
+	}
+	t.Fatalf("the server never reported a listening address within the deadline:\n%s", strings.Join(transcript, "\n--- poll ---\n"))
+	return ""
+}
+
+// validListeningAddress is parseListeningAddress plus the check that what it
+// found is actually an address.
+//
+// A read DRAINS, and the formatter wraps each drain in its own body, so a line
+// split across two reads arrives as two fragments with banner text wedged between
+// them and no concatenation can rejoin it. The formatter also terminates the
+// partial chunk, so the fragment looks like a whole line and the bare parse
+// returns a truncated address: the test would then fail on an unreachable port
+// and point the next reader at the network instead of at the parse. Rejecting
+// anything that is not host:port keeps polling instead.
+func validListeningAddress(output string) string {
+	addr := parseListeningAddress(output)
+	if addr == "" {
+		return ""
+	}
+	host, port, err := net.SplitHostPort(addr)
+	if err != nil || host == "" {
+		return ""
+	}
+	if number, err := strconv.Atoi(port); err != nil || number <= 0 {
+		return ""
+	}
+	return addr
 }
 
 func parseListeningAddress(output string) string {
@@ -427,7 +576,12 @@ func TestExecCommandReapsFinishedUnpolledSession(t *testing.T) {
 		t.Fatalf("session_id is not numeric: %v", err)
 	}
 
-	deadline := time.Now().Add(2 * time.Second)
+	// Ten seconds, not two. Inside this the helper must cold-start, run its 250ms
+	// sleep, exit and pass the reap, and a cold start alone was measured at up to
+	// 1.44s under load: two seconds was a budget with almost nothing left in it. The
+	// loop returns as soon as the session is gone, so the bound only costs anything
+	// when the reap genuinely never happens.
+	deadline := time.Now().Add(10 * time.Second)
 	for {
 		if _, ok := manager.Snapshot(sessionID); !ok {
 			return
@@ -739,6 +893,13 @@ func TestExecCommandTTYSessionAcceptsInputOnLinux(t *testing.T) {
 		t.Fatalf("session_id is not numeric: %v", err)
 	}
 
+	// SAME SHAPE AS THE SERVER TEST ABOVE, one window narrower.
+	//
+	// The shell is started with a 10ms yield, so it has almost certainly not
+	// reached `read` yet, and the echo then had exactly one fixed 1000ms window to
+	// travel back through the PTY. Writing early is fine, the terminal buffers it,
+	// but a loaded runner can spend that whole second still getting the shell up.
+	// So the echo is polled for rather than demanded within one window.
 	result := writeTool.Run(context.Background(), map[string]any{
 		"session_id":    sessionID,
 		"chars":         "hello\n",
@@ -747,11 +908,39 @@ func TestExecCommandTTYSessionAcceptsInputOnLinux(t *testing.T) {
 	if result.Status != StatusOK {
 		t.Fatalf("write_stdin status = %s: %s", result.Status, result.Output)
 	}
-	if !strings.Contains(result.Output, "got:hello") {
-		t.Fatalf("expected PTY input output, got %q", result.Output)
+	var transcript strings.Builder
+	transcript.WriteString(result.Output)
+	exitCode := result.Meta["exit_code"]
+	// BOTH THE ECHO AND THE EXIT, because the test asserts both.
+	//
+	// collect can return the drained echo on a yield timer before markDone has
+	// run, so a poll can carry got:hello with no exit_code meta yet. Stopping on
+	// the echo alone would then assert an exit that had not been observed. The
+	// condition deliberately keeps the echo term as well: gating on exit_code
+	// alone would stop polling before a slow shell start delivers the echo, which
+	// is the race this whole change exists to remove.
+	deadline := time.Now().Add(30 * time.Second)
+	for (!strings.Contains(transcript.String(), "got:hello") || exitCode == "") && time.Now().Before(deadline) {
+		poll := writeTool.Run(context.Background(), map[string]any{
+			"session_id":    sessionID,
+			"chars":         "",
+			"yield_time_ms": 250,
+		})
+		transcript.WriteString(poll.Output)
+		if code := poll.Meta["exit_code"]; code != "" {
+			exitCode = code
+		}
+		if poll.Status != StatusOK {
+			// The session is gone. Either the echo already arrived, which the loop
+			// condition will see, or it never will and the assertions below say so.
+			break
+		}
 	}
-	if result.Meta["exit_code"] != "0" {
-		t.Fatalf("expected exited session, got meta=%#v output=%q", result.Meta, result.Output)
+	if !strings.Contains(transcript.String(), "got:hello") {
+		t.Fatalf("expected PTY input output, got %q", transcript.String())
+	}
+	if exitCode != "0" {
+		t.Fatalf("expected exited session, got exit_code=%q output=%q", exitCode, transcript.String())
 	}
 }