Dining philosophers problem in Go fails unit test

I'm taking a Go course, that has an assignment as follows:

> Implement the dining philosopher's problem with the following
> constraints/modifications.
>
> - There should be 5 philosophers sharing chopsticks, with one chopstick between each adjacent pair of philosophers.
>
> - Each philosopher should eat only 3 times (not in an infinite loop as we did in lecture).
>
> - The philosophers pick up the chopsticks in any order, not lowest-numbered first (which we did in lecture).
>
> - In order to eat, a philosopher must get permission from a host which executes in its own goroutine.
>
> - The host allows no more than 2 philosophers to eat concurrently.
>
> Each philosopher is numbered, 1 through 5.
>
> When a philosopher starts eating (after it has obtained necessary
> locks) it prints "starting to eat <number>" on a line by itself, where
> <number> is the number of the philosopher.
>
> When a philosopher finishes eating (before it has released its locks)
> it prints "finishing eating <number>" on a line by itself, where
> <number> is the number of the philosopher.

My implementation:

package main

import (
	"fmt"
	"io"
	"math/rand"
	"os"
	"sync"
	"time"
)

const (
	NumPhilosophers    = 5
	NumEatMaxTimes     = 3
	NumMaxAllowedToEat = 2
)

type chopstick struct{ sync.Mutex }

type philosopher struct {
	num int
	cs  []*chopstick
}

func setTable() []*philosopher {
	cs := make([]*chopstick, NumPhilosophers)
	for i := 0; i < NumPhilosophers; i++ {
		cs[i] = new(chopstick)
	}
	ph := make([]*philosopher, NumPhilosophers)
	for i := 0; i < NumPhilosophers; i++ {
		ph[i] = &philosopher{i + 1, []*chopstick{cs[i], cs[(i+1)%NumPhilosophers]}}
	}

	return ph
}

func (ph philosopher) eat(sem chan int, wg *sync.WaitGroup, w io.Writer) {
	for i := 0; i < NumEatMaxTimes; i++ {
		/* Ask host for permission to eat */
		sem <- 1
		/*
			Pick any of the left or right chopsticks.
			Notice how the methods on the Mutex can be called directly on a chopstick due to embedding.
		*/
		firstCS := rand.Intn(2)
		secondCS := (firstCS + 1) % 2
		ph.cs[firstCS].Lock()
		ph.cs[secondCS].Lock()

		fmt.Fprintf(w, "Starting to eat %d\n", ph.num)
		x := rand.Intn(NumEatMaxTimes)
		time.Sleep(time.Duration(x) * time.Second)
		fmt.Fprintf(w, "Finishing eating %d\n", ph.num)

		ph.cs[secondCS].Unlock()
		ph.cs[firstCS].Unlock()
		<-sem
	}
	wg.Done()
}

func main() {
	run(os.Stdout)
}

func run(w io.Writer) {
	var sem = make(chan int, NumMaxAllowedToEat)
	rand.Seed(time.Now().UnixNano())
	var wg sync.WaitGroup

	allPh := setTable()
	wg.Add(len(allPh))
	for _, ph := range allPh {
		go ph.eat(sem, &wg, w)
	}
	wg.Wait()
}

Unit test:

func TestRun(t *testing.T) {
	var out bytes.Buffer
	run(&out)
	lines := strings.Split(strings.ReplaceAll(out.String(), "\r\n", "\n"), "\n")
	eating := make(map[int]bool)
	timesEaten := make(map[int]int)
	for _, line := range lines {
		if line == "" {
			continue
		}
		fmt.Println(line)
		tokens := strings.Fields(line)

		i, err := strconv.Atoi(tokens[len(tokens)-1])
		if err != nil {
			t.Errorf("Bad line: %s", line)
		}

		s := strings.ToLower(tokens[0])

		if s == "starting" {
			if len(eating) > (NumMaxAllowedToEat - 1) {
				t.Errorf("%v are eating at the same time", eating)
			}
			_, ok := eating[i]
			if ok {
				t.Errorf("%d started before finishing", i)
			}
			eating[i] = true
		} else if s == "finishing" {
			_, ok := eating[i]
			if !ok {
				t.Errorf("%d finished without starting", i)
			}

			delete(eating, i)

			timesEaten[i] = timesEaten[i] + 1
		}
	}

	for k, v := range timesEaten {
		if v > NumEatMaxTimes {
			t.Errorf("%d ate %d times", k, v)
		}
	}

	if len(timesEaten) != NumPhilosophers {
		t.Error("One or more didn't get to eat")
	}
}

The problem is, the test randomly fails. Below is one execution (line numbers added):

1. Starting to eat 5
2. Starting to eat 2
3. Finishing eating 2
4. Finishing eating 5
5. Starting to eat 3
6. Starting to eat 1
7. Finishing eating 1
8. Finishing eating 3
9. Starting to eat 2
10. Starting to eat 4
11. Finishing eating 4
12. Starting to eat 5
13. Finishing eating 2
14. Finishing eating 5
15. Starting to eat 3
16. Finishing eating 3
17. Starting to eat 1
18. Finishing eating 4
19. Finishing eating 1
20. Starting to eat 5
21. Finishing eating 5
22. Starting to eat 3
23. Finishing eating 3
24. Starting to eat 4
25. Starting to eat 2
26. Finishing eating 2
27. Starting to eat 1
28. Finishing eating 4
29. Finishing eating 1

--- FAIL: TestRun (12.01s)
    main_test.go:43: 4 finished without starting

Philosopher 4 has started on lines 10, and 24 and finished on lines 11, 18, and 28. Line 28 is unmatched, so the test correctly complains. However, I'm having a hard time finding the bug. Can you help?

Answering my own question, it turned out that byes.Buffer is not thread-safe. I ended up using go-fakeio library for the test as shown below.

s, err := fakeio.Stderr().Stdout().Do(run)
if err != nil {
    t.Errorf("%v", err)
}

The rest of the test remains the same. main.run function no longer needs an io.Writer since the fakeio library replaces stderr and stdout.