docs/architecture.md

@@ -1,10 +1,12 @@
 # Architecture
 
-> **Status:** the operator is built through Step 5 (health engine) of
+> **Status:** the operator is built through Step 6 (lease store) of
 > [docs/plans/2026-08-07-1747-proxy-operator.md](plans/2026-08-07-1747-proxy-operator.md).
 > This document currently covers the event/reconcile flow; the components
-> table and the Decisions section arrive with Step 10, and the diagrams
+> table and the Decisions section arrive with Step 10. The lease store
-> below grow as the lease store, discovery API, and orphan GC land.
+> (`internal/lease/`) is HTTP-driven, not cluster-event-driven, so its
+> diagram lands together with the discovery API in Step 7; the orphan-GC
+> flow lands with Step 9.
 
 ## Event flow: cluster events → reconciler functions
 

docs/plans-executions/2026-08-07-1747-proxy-operator.md

@@ -10,7 +10,7 @@ Pairs with [docs/plans/2026-08-07-1747-proxy-operator.md](../plans/2026-08-07-17
 - [x] Step 3 — Kubernetes pod provider (`internal/provider/kubernetes/`; first built as an in-memory mock, then replaced — see the two Step 3 sections below)
 - [x] Step 4 — Reconciler (`internal/controller/`)
 - [x] Step 5 — Health engine (`internal/health/`)
-- [ ] Step 6 — Lease store (`internal/lease/`)
+- [x] Step 6 — Lease store (`internal/lease/`)
 - [ ] Step 7 — Discovery API (`internal/discovery/`)
 - [ ] Step 8 — GCP provider (`internal/provider/gcp/`)
 - [ ] Step 9 — Orphan GC + metrics
@@ -694,3 +694,67 @@ engine deliberately knows nothing about conditions except reading one at
 seed time, keeping the state/representation split honest. The
 `hint`-driven `wg.Go` idiom (Go 1.25+) replaced the classic
 `wg.Add/defer wg.Done` in the worker pool.
+
+## Step 6 — Lease store (`internal/lease/`)
+
+Implemented `store.go` per the plan: `Acquire` takes the whole candidate
+set so selection and insertion happen under the one store mutex (no
+overcommit between concurrent requests), selection is a linear scan +
+`slices.SortFunc` on `(activeLeases asc, latency asc, name asc)`,
+`AcquireStats{Considered, AtCapacity, InCooldown}` feeds Step 7's 409
+body, cooldowns live in a `map[{proxy, target}]time.Time` (empty target =
+global pool), and expired leases are retained for `CooldownWindow` past
+their TTL so a late `Report` — arriving exactly when a proxy is being
+rate-limited — still resolves and records its cooldown.
+
+Semantics pinned against the spec (§8) rather than guessed:
+
+- Report results are exactly `ok | rate_limited | banned` (`ParseResult`
+  gives the API layer its 400 check). `rate_limited` and `banned` both
+  record a cooldown for the same window; `ok` records nothing.
+  Distinguishing ban duration from rate-limit duration would be a second
+  knob the spec doesn't ask for — noted for the Decisions section.
+- Cooldown scoping: the global cooldown (empty target) always applies; a
+  target-scoped cooldown additionally blocks acquisitions for that target;
+  acquisitions without a target see only the global pool ("a proxy
+  rate-limited by one site is still fine for everyone else").
+- A `Report` without a target falls back to the lease's own target before
+  falling back to global — so a client that leased with a target doesn't
+  accidentally poison the whole proxy by omitting it in the report.
+
+Design notes:
+
+- **Correctness never depends on the sweep.** Every read path
+  (`Acquire`/`ActiveCount`/`Counts`) compares `ExpiresAt` against the
+  injected clock, so TTL expiry frees capacity immediately even if the
+  background loop hasn't run; the sweep is purely garbage collection. The
+  plan's `ExpireLoop` became `Start(ctx)` + `NeedLeaderElection() false`
+  so the store satisfies `manager.Runnable` directly — Step 10 just
+  `mgr.Add(store)`s it. Not leader-elected because lease state is
+  per-process and must expire wherever the discovery API is serving.
+- The store knows nothing about Proxy objects — `Candidate` carries the
+  opaque key, `MaxLeases`, and latency; the discovery layer does the
+  health/attribute filtering. The spec's `LeaseStore` interface will be
+  defined consumer-side in `internal/discovery` (Step 7), per Go idiom;
+  this package exports only the concrete in-memory `*Store`.
+- Lease IDs come from `crypto/rand.Text()` (Go 1.24+); returned `Lease`
+  values are copies so callers can't mutate store internals.
+
+Tests (94.8% coverage, `-race -count=2` clean): capacity + release
+freeing slots, `MaxLeases=0` unleasable, least-loaded/latency/name
+selection order, target-scoped vs global cooldown scoping, cooldown
+expiry via the injected fake clock, TTL freeing capacity with no sweep,
+report-on-expired-but-retained lease (then `ErrUnknownLease` after
+retention), `ok` recording nothing, idempotent release, `ParseResult`,
+40 concurrent acquires against `MaxLeases=5` granting exactly 5, and the
+`Start` loop sweeping then stopping cleanly on cancel.
+
+```bash
+go test -race -count=2 ./internal/lease/
+make test    # whole repo green, other packages' coverage unchanged
+```
+
+Worth noting: `docs/architecture.md` was not extended this step — the
+lease store is HTTP-driven, not cluster-event-driven, so its diagram
+belongs with the discovery API and lands in Step 7 (banner updated to say
+so).

internal/lease/store.go

@@ -0,0 +1,319 @@
+// Package lease implements the in-memory lease store behind the discovery
+// API: TTL-based proxy assignment with server-side usage tracking and
+// per-(proxy, target) cooldowns. Accepted prototype limitation, documented
+// in the README: state is per-process, so an operator restart drops all
+// leases and cooldowns — clients must tolerate a lease vanishing (their
+// requests still work; they just re-lease).
+package lease
+
+import (
+	"cmp"
+	"context"
+	"crypto/rand"
+	"errors"
+	"fmt"
+	"slices"
+	"strings"
+	"sync"
+	"time"
+)
+
+// Result is a client's report of how a leased proxy behaved against a
+// target. ResultRateLimited and ResultBanned record a cooldown; ResultOK is
+// an acknowledgement and records nothing.
+type Result string
+
+const (
+	ResultOK          Result = "ok"
+	ResultRateLimited Result = "rate_limited"
+	ResultBanned      Result = "banned"
+)
+
+// ParseResult maps a wire value to a Result; ok is false for anything
+// unknown, which the API layer turns into a 400.
+func ParseResult(s string) (Result, bool) {
+	switch r := Result(s); r {
+	case ResultOK, ResultRateLimited, ResultBanned:
+		return r, true
+	default:
+		return "", false
+	}
+}
+
+var (
+	// ErrNoMatch means no candidate could take a lease; AcquireStats says
+	// why, and the API layer turns both into the 409 body.
+	ErrNoMatch = errors.New("lease: no candidate available")
+	// ErrUnknownLease means the lease ID does not resolve (404). Reports on
+	// recently expired leases do NOT hit this — see the retention note on
+	// Store.
+	ErrUnknownLease = errors.New("lease: unknown lease id")
+)
+
+// Candidate is one leasable proxy as seen by the caller at selection time.
+// The store itself knows nothing about Proxy objects — the discovery layer
+// filters for health/attributes and passes what selection needs.
+type Candidate struct {
+	// Proxy is the opaque proxy key ("namespace/name").
+	Proxy string
+	// MaxLeases caps concurrent leases; 0 means unleasable.
+	MaxLeases int32
+	// Latency is the proxy's last reported latency, used as the tie-break.
+	Latency time.Duration
+}
+
+// Lease is a granted assignment. Values returned by the store are copies;
+// mutating them does not affect the store.
+type Lease struct {
+	ID        string
+	Proxy     string
+	Target    string
+	ExpiresAt time.Time
+}
+
+// AcquireRequest carries the candidate set and lease parameters. Acquire
+// deliberately takes the whole candidate set, not a pre-chosen proxy:
+// selection and insertion must happen under one lock, or two concurrent
+// requests both see "3 of 5 used" and overcommit.
+type AcquireRequest struct {
+	Candidates []Candidate
+	// Target scopes the cooldown check; empty means the global pool.
+	Target string
+	TTL    time.Duration
+}
+
+// AcquireStats explains an ErrNoMatch (and is returned on success too):
+// every candidate is either leased, at capacity, or in cooldown.
+type AcquireStats struct {
+	Considered int
+	AtCapacity int
+	InCooldown int
+}
+
+type cooldownKey struct{ proxy, target string }
+
+// Store is the in-memory lease store. One mutex guards everything: at tens
+// of proxies and human-rate QPS, sharding would be premature complexity.
+//
+// Retention: an expired lease is kept for CooldownWindow past its TTL so a
+// Report arriving just after expiry still resolves — which matters most
+// exactly when a proxy is being rate-limited. Acquire and the counts ignore
+// retained leases; only the sweep finally drops them.
+type Store struct {
+	// CooldownWindow is how long a reported proxy/target pair is excluded
+	// from selection (default 15m; --lease-cooldown in Step 10).
+	CooldownWindow time.Duration
+	// SweepInterval is how often the expiry sweep runs (default 30s).
+	SweepInterval time.Duration
+
+	now func() time.Time
+
+	mu        sync.Mutex
+	byID      map[string]*Lease
+	byProxy   map[string]map[string]*Lease
+	cooldowns map[cooldownKey]time.Time
+}
+
+// NewStore returns a ready Store. A non-positive cooldownWindow selects the
+// 15-minute default.
+func NewStore(cooldownWindow time.Duration) *Store {
+	if cooldownWindow <= 0 {
+		cooldownWindow = 15 * time.Minute
+	}
+	return &Store{
+		CooldownWindow: cooldownWindow,
+		SweepInterval:  30 * time.Second,
+		now:            time.Now,
+		byID:           map[string]*Lease{},
+		byProxy:        map[string]map[string]*Lease{},
+		cooldowns:      map[cooldownKey]time.Time{},
+	}
+}
+
+// Acquire selects the least-loaded eligible candidate (ties: lowest
+// latency, then name, so selection is deterministic and testable) and
+// grants a lease on it.
+func (s *Store) Acquire(_ context.Context, req AcquireRequest) (*Lease, AcquireStats, error) {
+	stats := AcquireStats{Considered: len(req.Candidates)}
+	if req.TTL <= 0 {
+		return nil, stats, fmt.Errorf("lease: non-positive TTL %v", req.TTL)
+	}
+	now := s.now()
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	type eligible struct {
+		cand   Candidate
+		active int
+	}
+	var elig []eligible
+	for _, c := range req.Candidates {
+		if s.inCooldownLocked(c.Proxy, req.Target, now) {
+			stats.InCooldown++
+			continue
+		}
+		active := s.activeCountLocked(c.Proxy, now)
+		if int32(active) >= c.MaxLeases {
+			stats.AtCapacity++
+			continue
+		}
+		elig = append(elig, eligible{cand: c, active: active})
+	}
+	if len(elig) == 0 {
+		return nil, stats, ErrNoMatch
+	}
+
+	slices.SortFunc(elig, func(a, b eligible) int {
+		if c := cmp.Compare(a.active, b.active); c != 0 {
+			return c
+		}
+		if c := cmp.Compare(a.cand.Latency, b.cand.Latency); c != 0 {
+			return c
+		}
+		return strings.Compare(a.cand.Proxy, b.cand.Proxy)
+	})
+
+	l := &Lease{
+		ID:        rand.Text(),
+		Proxy:     elig[0].cand.Proxy,
+		Target:    req.Target,
+		ExpiresAt: now.Add(req.TTL),
+	}
+	s.byID[l.ID] = l
+	if s.byProxy[l.Proxy] == nil {
+		s.byProxy[l.Proxy] = map[string]*Lease{}
+	}
+	s.byProxy[l.Proxy][l.ID] = l
+
+	granted := *l
+	return &granted, stats, nil
+}
+
+// Release drops a lease early. Idempotent: releasing an unknown or already
+// expired lease is a no-op, so the API's DELETE can always answer 204.
+func (s *Store) Release(_ context.Context, id string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.dropLocked(id)
+}
+
+// Report records the outcome of using a lease. Rate-limited and banned
+// results put the (proxy, target) pair in cooldown — target taken from the
+// report, falling back to the lease's own target, falling back to the
+// global pool. Reports on recently expired leases still resolve (see the
+// retention note on Store).
+func (s *Store) Report(_ context.Context, id string, result Result, target string) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	l, ok := s.byID[id]
+	if !ok {
+		return ErrUnknownLease
+	}
+	if result == ResultOK {
+		return nil
+	}
+	if target == "" {
+		target = l.Target
+	}
+	s.cooldowns[cooldownKey{proxy: l.Proxy, target: target}] = s.now().Add(s.CooldownWindow)
+	return nil
+}
+
+// ActiveCount returns the number of unexpired leases held on one proxy.
+func (s *Store) ActiveCount(proxy string) int {
+	now := s.now()
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.activeCountLocked(proxy, now)
+}
+
+// Counts returns the active-lease count per proxy, for the discovery list
+// endpoint and the metrics collector. Proxies with no active leases are
+// absent from the map.
+func (s *Store) Counts() map[string]int {
+	now := s.now()
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	counts := make(map[string]int, len(s.byProxy))
+	for proxy := range s.byProxy {
+		if n := s.activeCountLocked(proxy, now); n > 0 {
+			counts[proxy] = n
+		}
+	}
+	return counts
+}
+
+// Start runs the expiry sweep until ctx ends; it satisfies
+// manager.Runnable so cmd/main.go can mgr.Add the store directly.
+func (s *Store) Start(ctx context.Context) error {
+	ticker := time.NewTicker(s.SweepInterval)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return nil
+		case <-ticker.C:
+			s.sweep(s.now())
+		}
+	}
+}
+
+// NeedLeaderElection is false: lease state is per-process and the discovery
+// API serves wherever this process runs, so the sweep must run there too.
+func (s *Store) NeedLeaderElection() bool { return false }
+
+// sweep drops leases past their retention window and elapsed cooldowns.
+// Correctness never depends on sweep timing — every read path checks
+// expiry against the clock — so this is purely garbage collection.
+func (s *Store) sweep(now time.Time) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for id, l := range s.byID {
+		if now.After(l.ExpiresAt.Add(s.CooldownWindow)) {
+			s.dropLocked(id)
+		}
+	}
+	for k, until := range s.cooldowns {
+		if now.After(until) {
+			delete(s.cooldowns, k)
+		}
+	}
+}
+
+func (s *Store) dropLocked(id string) {
+	l, ok := s.byID[id]
+	if !ok {
+		return
+	}
+	delete(s.byID, id)
+	delete(s.byProxy[l.Proxy], id)
+	if len(s.byProxy[l.Proxy]) == 0 {
+		delete(s.byProxy, l.Proxy)
+	}
+}
+
+func (s *Store) activeCountLocked(proxy string, now time.Time) int {
+	n := 0
+	for _, l := range s.byProxy[proxy] {
+		if now.Before(l.ExpiresAt) {
+			n++
+		}
+	}
+	return n
+}
+
+// inCooldownLocked: the global cooldown (empty target) always applies; a
+// target-scoped cooldown additionally applies to acquisitions for that
+// target. An acquisition without a target sees only the global pool — a
+// proxy rate-limited by one site is still fine for everyone else.
+func (s *Store) inCooldownLocked(proxy, target string, now time.Time) bool {
+	if until, ok := s.cooldowns[cooldownKey{proxy: proxy}]; ok && now.Before(until) {
+		return true
+	}
+	if target == "" {
+		return false
+	}
+	until, ok := s.cooldowns[cooldownKey{proxy: proxy, target: target}]
+	return ok && now.Before(until)
+}

internal/lease/store_test.go

@@ -0,0 +1,365 @@
+package lease
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+	"time"
+)
+
+// fakeClock is an injectable, manually advanced clock.
+type fakeClock struct {
+	mu  sync.Mutex
+	cur time.Time
+}
+
+func newFakeClock() *fakeClock {
+	return &fakeClock{cur: time.Date(2026, 8, 9, 12, 0, 0, 0, time.UTC)}
+}
+
+func (c *fakeClock) Now() time.Time {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return c.cur
+}
+
+func (c *fakeClock) Advance(d time.Duration) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	c.cur = c.cur.Add(d)
+}
+
+func newTestStore() (*Store, *fakeClock) {
+	s := NewStore(15 * time.Minute)
+	clock := newFakeClock()
+	s.now = clock.Now
+	return s, clock
+}
+
+func candidate(proxy string, maxLeases int32, latency time.Duration) Candidate {
+	return Candidate{Proxy: proxy, MaxLeases: maxLeases, Latency: latency}
+}
+
+func mustAcquire(t *testing.T, s *Store, req AcquireRequest) *Lease {
+	t.Helper()
+	l, _, err := s.Acquire(context.Background(), req)
+	if err != nil {
+		t.Fatalf("Acquire: %v", err)
+	}
+	return l
+}
+
+func TestAcquire_capacity(t *testing.T) {
+	t.Parallel()
+	s, _ := newTestStore()
+	req := AcquireRequest{Candidates: []Candidate{candidate("ns/p1", 2, 0)}, TTL: time.Minute}
+
+	l1 := mustAcquire(t, s, req)
+	l2 := mustAcquire(t, s, req)
+	if l1.ID == l2.ID {
+		t.Fatal("two leases share an ID")
+	}
+	if got := s.ActiveCount("ns/p1"); got != 2 {
+		t.Fatalf("ActiveCount = %d, want 2", got)
+	}
+
+	_, stats, err := s.Acquire(context.Background(), req)
+	if !errors.Is(err, ErrNoMatch) {
+		t.Fatalf("third acquire error = %v, want ErrNoMatch", err)
+	}
+	want := AcquireStats{Considered: 1, AtCapacity: 1}
+	if stats != want {
+		t.Errorf("stats = %+v, want %+v", stats, want)
+	}
+
+	// Early release frees the slot again.
+	s.Release(context.Background(), l1.ID)
+	mustAcquire(t, s, req)
+}
+
+func TestAcquire_maxLeasesZeroIsUnleasable(t *testing.T) {
+	t.Parallel()
+	s, _ := newTestStore()
+	_, stats, err := s.Acquire(context.Background(), AcquireRequest{
+		Candidates: []Candidate{candidate("ns/p1", 0, 0)},
+		TTL:        time.Minute,
+	})
+	if !errors.Is(err, ErrNoMatch) {
+		t.Fatalf("err = %v, want ErrNoMatch", err)
+	}
+	if stats.AtCapacity != 1 {
+		t.Errorf("stats = %+v, want the unleasable proxy counted AtCapacity", stats)
+	}
+}
+
+func TestAcquire_selectionOrder(t *testing.T) {
+	t.Parallel()
+
+	t.Run("least loaded wins", func(t *testing.T) {
+		t.Parallel()
+		s, _ := newTestStore()
+		mustAcquire(t, s, AcquireRequest{
+			Candidates: []Candidate{candidate("ns/a", 5, 10*time.Millisecond)}, TTL: time.Minute,
+		})
+		l := mustAcquire(t, s, AcquireRequest{
+			Candidates: []Candidate{
+				candidate("ns/a", 5, 10*time.Millisecond), // 1 active, lower latency
+				candidate("ns/b", 5, 90*time.Millisecond), // 0 active
+			},
+			TTL: time.Minute,
+		})
+		if l.Proxy != "ns/b" {
+			t.Errorf("chose %s, want the least-loaded ns/b", l.Proxy)
+		}
+	})
+
+	t.Run("latency breaks the load tie", func(t *testing.T) {
+		t.Parallel()
+		s, _ := newTestStore()
+		l := mustAcquire(t, s, AcquireRequest{
+			Candidates: []Candidate{
+				candidate("ns/a", 5, 90*time.Millisecond),
+				candidate("ns/b", 5, 10*time.Millisecond),
+			},
+			TTL: time.Minute,
+		})
+		if l.Proxy != "ns/b" {
+			t.Errorf("chose %s, want the lower-latency ns/b", l.Proxy)
+		}
+	})
+
+	t.Run("name breaks a full tie deterministically", func(t *testing.T) {
+		t.Parallel()
+		s, _ := newTestStore()
+		l := mustAcquire(t, s, AcquireRequest{
+			Candidates: []Candidate{
+				candidate("ns/b", 5, 10*time.Millisecond),
+				candidate("ns/a", 5, 10*time.Millisecond),
+			},
+			TTL: time.Minute,
+		})
+		if l.Proxy != "ns/a" {
+			t.Errorf("chose %s, want ns/a (lexicographic tie-break)", l.Proxy)
+		}
+	})
+}
+
+func TestAcquire_cooldownScoping(t *testing.T) {
+	t.Parallel()
+	s, _ := newTestStore()
+	cands := []Candidate{candidate("ns/p1", 5, 0)}
+
+	l := mustAcquire(t, s, AcquireRequest{Candidates: cands, Target: "example.com", TTL: time.Minute})
+	if err := s.Report(context.Background(), l.ID, ResultRateLimited, "example.com"); err != nil {
+		t.Fatalf("Report: %v", err)
+	}
+
+	// Same target: excluded.
+	_, stats, err := s.Acquire(context.Background(), AcquireRequest{
+		Candidates: cands, Target: "example.com", TTL: time.Minute,
+	})
+	if !errors.Is(err, ErrNoMatch) || stats.InCooldown != 1 {
+		t.Errorf("same-target acquire = (%v, %+v), want ErrNoMatch with InCooldown=1", err, stats)
+	}
+
+	// Different target: fine.
+	mustAcquire(t, s, AcquireRequest{Candidates: cands, Target: "other.org", TTL: time.Minute})
+
+	// No target (global pool): a target-scoped cooldown does not apply.
+	mustAcquire(t, s, AcquireRequest{Candidates: cands, TTL: time.Minute})
+}
+
+func TestAcquire_globalCooldownBlocksEverything(t *testing.T) {
+	t.Parallel()
+	s, _ := newTestStore()
+	cands := []Candidate{candidate("ns/p1", 5, 0)}
+
+	// A lease without a target, reported banned without a target: the
+	// cooldown lands on the global pool.
+	l := mustAcquire(t, s, AcquireRequest{Candidates: cands, TTL: time.Minute})
+	if err := s.Report(context.Background(), l.ID, ResultBanned, ""); err != nil {
+		t.Fatalf("Report: %v", err)
+	}
+
+	for _, target := range []string{"", "example.com"} {
+		_, stats, err := s.Acquire(context.Background(), AcquireRequest{
+			Candidates: cands, Target: target, TTL: time.Minute,
+		})
+		if !errors.Is(err, ErrNoMatch) || stats.InCooldown != 1 {
+			t.Errorf("acquire(target=%q) = (%v, %+v), want global cooldown to block", target, err, stats)
+		}
+	}
+}
+
+func TestAcquire_cooldownExpires(t *testing.T) {
+	t.Parallel()
+	s, clock := newTestStore()
+	cands := []Candidate{candidate("ns/p1", 5, 0)}
+
+	l := mustAcquire(t, s, AcquireRequest{Candidates: cands, TTL: time.Minute})
+	if err := s.Report(context.Background(), l.ID, ResultRateLimited, ""); err != nil {
+		t.Fatalf("Report: %v", err)
+	}
+	if _, _, err := s.Acquire(context.Background(), AcquireRequest{Candidates: cands, TTL: time.Minute}); !errors.Is(err, ErrNoMatch) {
+		t.Fatal("expected cooldown to block immediately after the report")
+	}
+
+	clock.Advance(15*time.Minute + time.Second)
+	mustAcquire(t, s, AcquireRequest{Candidates: cands, TTL: time.Minute})
+}
+
+func TestExpiry_freesCapacityWithoutSweep(t *testing.T) {
+	t.Parallel()
+	s, clock := newTestStore()
+	req := AcquireRequest{Candidates: []Candidate{candidate("ns/p1", 1, 0)}, TTL: time.Minute}
+
+	mustAcquire(t, s, req)
+	if _, _, err := s.Acquire(context.Background(), req); !errors.Is(err, ErrNoMatch) {
+		t.Fatal("capacity 1 not enforced")
+	}
+
+	clock.Advance(2 * time.Minute)
+	// No sweep has run; expiry must still free capacity and zero the counts.
+	if got := s.ActiveCount("ns/p1"); got != 0 {
+		t.Fatalf("ActiveCount after TTL = %d, want 0", got)
+	}
+	if counts := s.Counts(); len(counts) != 0 {
+		t.Fatalf("Counts after TTL = %v, want empty", counts)
+	}
+	mustAcquire(t, s, req)
+}
+
+func TestReport_expiredButRetainedLease(t *testing.T) {
+	t.Parallel()
+	s, clock := newTestStore()
+	cands := []Candidate{candidate("ns/p1", 5, 0)}
+
+	l := mustAcquire(t, s, AcquireRequest{Candidates: cands, Target: "example.com", TTL: time.Minute})
+
+	// TTL lapses; the report arrives late — exactly when the proxy is being
+	// rate-limited, which is when the cooldown matters most.
+	clock.Advance(5 * time.Minute)
+	s.sweep(clock.Now())
+	if err := s.Report(context.Background(), l.ID, ResultRateLimited, ""); err != nil {
+		t.Fatalf("Report on an expired-but-retained lease: %v", err)
+	}
+	// The cooldown fell back to the lease's own target.
+	_, stats, err := s.Acquire(context.Background(), AcquireRequest{
+		Candidates: cands, Target: "example.com", TTL: time.Minute,
+	})
+	if !errors.Is(err, ErrNoMatch) || stats.InCooldown != 1 {
+		t.Errorf("acquire = (%v, %+v), want cooldown from the late report", err, stats)
+	}
+
+	// Past the retention window the sweep finally drops it.
+	clock.Advance(15 * time.Minute)
+	s.sweep(clock.Now())
+	if err := s.Report(context.Background(), l.ID, ResultRateLimited, ""); !errors.Is(err, ErrUnknownLease) {
+		t.Errorf("Report after retention = %v, want ErrUnknownLease", err)
+	}
+}
+
+func TestReport_okRecordsNothing(t *testing.T) {
+	t.Parallel()
+	s, _ := newTestStore()
+	cands := []Candidate{candidate("ns/p1", 5, 0)}
+	l := mustAcquire(t, s, AcquireRequest{Candidates: cands, Target: "example.com", TTL: time.Minute})
+
+	if err := s.Report(context.Background(), l.ID, ResultOK, "example.com"); err != nil {
+		t.Fatalf("Report(ok): %v", err)
+	}
+	mustAcquire(t, s, AcquireRequest{Candidates: cands, Target: "example.com", TTL: time.Minute})
+}
+
+func TestRelease_isIdempotent(t *testing.T) {
+	t.Parallel()
+	s, _ := newTestStore()
+	l := mustAcquire(t, s, AcquireRequest{Candidates: []Candidate{candidate("ns/p1", 1, 0)}, TTL: time.Minute})
+
+	s.Release(context.Background(), l.ID)
+	s.Release(context.Background(), l.ID)
+	s.Release(context.Background(), "never-existed")
+	if got := s.ActiveCount("ns/p1"); got != 0 {
+		t.Errorf("ActiveCount = %d, want 0", got)
+	}
+}
+
+func TestParseResult(t *testing.T) {
+	t.Parallel()
+	for _, valid := range []string{"ok", "rate_limited", "banned"} {
+		if _, ok := ParseResult(valid); !ok {
+			t.Errorf("ParseResult(%q) rejected a valid value", valid)
+		}
+	}
+	for _, invalid := range []string{"", "OK", "throttled", "rate-limited"} {
+		if _, ok := ParseResult(invalid); ok {
+			t.Errorf("ParseResult(%q) accepted an invalid value", invalid)
+		}
+	}
+}
+
+func TestAcquire_concurrentNeverOvercommits(t *testing.T) {
+	t.Parallel()
+	s, _ := newTestStore()
+	req := AcquireRequest{Candidates: []Candidate{candidate("ns/p1", 5, 0)}, TTL: time.Minute}
+
+	const attempts = 40
+	var wg sync.WaitGroup
+	granted := make(chan *Lease, attempts)
+	for range attempts {
+		wg.Go(func() {
+			if l, _, err := s.Acquire(context.Background(), req); err == nil {
+				granted <- l
+			}
+		})
+	}
+	wg.Wait()
+	close(granted)
+
+	var n int
+	for range granted {
+		n++
+	}
+	if n != 5 {
+		t.Errorf("%d of %d concurrent acquires granted, want exactly MaxLeases=5", n, attempts)
+	}
+	if got := s.ActiveCount("ns/p1"); got != 5 {
+		t.Errorf("ActiveCount = %d, want 5", got)
+	}
+}
+
+func TestStart_sweepsAndStops(t *testing.T) {
+	t.Parallel()
+	s, clock := newTestStore()
+	s.SweepInterval = time.Millisecond
+
+	l := mustAcquire(t, s, AcquireRequest{Candidates: []Candidate{candidate("ns/p1", 5, 0)}, TTL: time.Minute})
+	clock.Advance(20 * time.Minute) // past TTL + retention
+
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan error, 1)
+	go func() { done <- s.Start(ctx) }()
+
+	deadline := time.After(5 * time.Second)
+	for {
+		if err := s.Report(context.Background(), l.ID, ResultOK, ""); errors.Is(err, ErrUnknownLease) {
+			break
+		}
+		select {
+		case <-deadline:
+			t.Fatal("sweep never dropped the lease")
+		case <-time.After(5 * time.Millisecond):
+		}
+	}
+
+	cancel()
+	select {
+	case err := <-done:
+		if err != nil {
+			t.Errorf("Start returned %v, want nil", err)
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("Start did not stop on cancel")
+	}
+}