diff --git a/go/internal/ingest/board_webhook.go b/go/internal/ingest/board_webhook.go
index cc8bd5626..c92aa2dfd 100644
--- a/go/internal/ingest/board_webhook.go
+++ b/go/internal/ingest/board_webhook.go
@@ -17,6 +17,7 @@ import (
 	"log/slog"
 	"strings"
 	"sync/atomic"
+	"time"
 
 	"github.com/RigelBuild/compass/go/internal/forge"
 	compassv1internal "github.com/RigelBuild/compass/go/internal/gen/compass/v1"
@@ -30,10 +31,15 @@ var boardWebhookDrops = expvar.NewInt("compass_board_webhook_drops")
 
 // defaultBoardQueueSize is the bounded drain-queue depth. Sized to absorb a
 // normal edit-storm burst between drains; a sustained overflow (the drain paused
-// on ErrBudgetExhausted while events keep arriving) drops with the metric+Warn
-// and is healed by the T3 reconciler.
+// on ErrBudgetExhausted while events keep arriving) drops with the metric+Warn.
+// The sweep heals a dropped issue; a dropped CI-only PR change waits for the PR's
+// next event.
 const defaultBoardQueueSize = 1024
 
+// defaultBoardRetryWait re-drains held PR keys when the budget error gave no
+// reset hint.
+const defaultBoardRetryWait = time.Minute
+
 // issueHydrator is the conditional point-read seam (satisfied structurally by
 // *forge.GitHub via GetIssueConditional, notify_reader.go:129). Defined locally
 // so this package imports no concrete forge client on the hydrate path.
@@ -82,6 +88,12 @@ type BoardWebhookArm struct {
 	targets  TargetChecker
 	log      *slog.Logger
 	dropped  atomic.Int64
+
+	// pending holds PR keys a budget pause left un-hydrated. The sweep cannot
+	// heal a CI-only change, so the drain retries them. Drain goroutine only.
+	pending    map[boardCoord]struct{}
+	maxPending int
+	after      func(time.Duration) <-chan time.Time
 }
 
 // NewBoardWebhookArm returns an arm that hydrates each accepted event through h,
@@ -103,6 +115,10 @@ func NewBoardWebhookArm(h issueHydrator, ing *Ingester, targets TargetChecker, c
 		numbers:  cfg.PullNumbers,
 		targets:  targets,
 		log:      log,
+
+		pending:    map[boardCoord]struct{}{},
+		maxPending: size,
+		after:      time.After,
 	}
 }
 
@@ -130,18 +146,38 @@ func (a *BoardWebhookArm) Enqueue(_ context.Context, ev forge.ForgeEvent) {
 
 // Run drains the queue until ctx is cancelled — then it returns nil (clean
 // shutdown, driver.go:95-99 idiom). Each drain COALESCES per coordinate before
-// hydrating, so an N-event burst on one issue costs one GET.
+// hydrating, so an N-event burst on one issue costs one GET. After a budget
+// pause it re-drains the held PR keys once the gate's reset hint passes.
 func (a *BoardWebhookArm) Run(ctx context.Context) error {
+	var retry <-chan time.Time
 	for {
 		select {
 		case <-ctx.Done():
 			return nil
 		case ev := <-a.queue:
-			a.drainBatch(ctx, ev)
+			if wait, paused := a.drainBatch(ctx, ev); paused && retry == nil {
+				retry = a.retryTimer(wait)
+			}
+		case <-retry:
+			retry = nil
+			if wait, paused := a.retryPending(ctx); paused {
+				retry = a.retryTimer(wait)
+			}
 		}
 	}
 }
 
+// retryTimer arms the re-drain for held PR keys; nil when none are held.
+func (a *BoardWebhookArm) retryTimer(wait time.Duration) <-chan time.Time {
+	if len(a.pending) == 0 {
+		return nil
+	}
+	if wait <= 0 {
+		wait = defaultBoardRetryWait
+	}
+	return a.after(wait)
+}
+
 // boardRelevant reports whether an event changes what the board shows. Issue
 // events count on OPENED, STATE or UPDATE; an issue COMMENT would burn one
 // hydrate per comment for nothing shown. Every PR event counts, since reviews
@@ -173,11 +209,12 @@ func (a *BoardWebhookArm) boardRelevant(ev forge.ForgeEvent) bool {
 // drainBatch coalesces the event that woke the drain and every other queued
 // event into distinct keys: normalized repo, kind and number, or head SHA for a
 // CHECKS event with no number. An edit storm and a mixed-case duplicate both
-// collapse to one key. It then hydrates + sinks each key once, in arrival order.
-func (a *BoardWebhookArm) drainBatch(ctx context.Context, first forge.ForgeEvent) {
+// collapse to one key. A key already held for retry is left to the timer, so a
+// held PR whose budget is still out never pauses a fresh batch. It then hydrates
+// + sinks each key once, in arrival order, and reports a pause.
+func (a *BoardWebhookArm) drainBatch(ctx context.Context, first forge.ForgeEvent) (time.Duration, bool) {
 	seen := map[boardCoord]struct{}{}
 	var order []boardCoord
-
 	add := func(ev forge.ForgeEvent) {
 		c := boardCoord{repo: normalizeBoardRepo(ev.Repo), kind: ev.Kind, number: ev.Number}
 		if ev.Number == 0 {
@@ -186,6 +223,9 @@ func (a *BoardWebhookArm) drainBatch(ctx context.Context, first forge.ForgeEvent
 		if _, ok := seen[c]; ok {
 			return
 		}
+		if _, held := a.pending[c]; held {
+			return
+		}
 		seen[c] = struct{}{}
 		order = append(order, c)
 	}
@@ -196,29 +236,69 @@ func (a *BoardWebhookArm) drainBatch(ctx context.Context, first forge.ForgeEvent
 		case ev := <-a.queue:
 			add(ev)
 		default:
-			goto process
+			return a.process(ctx, order, seen)
 		}
 	}
+}
 
-process:
-	for _, c := range order {
+// retryPending re-drains the PR keys held by an earlier budget pause.
+func (a *BoardWebhookArm) retryPending(ctx context.Context) (time.Duration, bool) {
+	order, seen := a.takePending()
+	return a.process(ctx, order, seen)
+}
+
+// takePending empties the held PR keys into a fresh batch.
+func (a *BoardWebhookArm) takePending() ([]boardCoord, map[boardCoord]struct{}) {
+	seen := make(map[boardCoord]struct{}, len(a.pending))
+	order := make([]boardCoord, 0, len(a.pending))
+	for c := range a.pending {
+		seen[c] = struct{}{}
+		order = append(order, c)
+	}
+	clear(a.pending)
+	return order, seen
+}
+
+// hold keeps the batch's un-hydrated PR keys for a retry; issue keys are left to
+// the sweep, which sees their updated_at move. Past maxPending a key is dropped.
+func (a *BoardWebhookArm) hold(rest []boardCoord) {
+	for _, c := range rest {
+		if c.kind != compassv1internal.ForgeArtifactKind_FORGE_ARTIFACT_KIND_PULL_REQUEST {
+			continue
+		}
+		if _, ok := a.pending[c]; !ok && len(a.pending) >= a.maxPending {
+			a.dropped.Add(1)
+			boardWebhookDrops.Add(1)
+			continue
+		}
+		a.pending[c] = struct{}{}
+	}
+}
+
+// process hydrates each key once. A budget error holds the remaining PR keys
+// and returns the gate's reset hint.
+func (a *BoardWebhookArm) process(ctx context.Context, order []boardCoord, seen map[boardCoord]struct{}) (time.Duration, bool) {
+	for i, c := range order {
 		if err := a.resolveAndSink(ctx, c, seen); err != nil {
 			if errors.Is(err, forge.ErrBudgetExhausted) {
-				// Budget exhausted pauses the drain: abandon the rest of this
-				// batch (the reconciler heals the un-hydrated coordinates) and
-				// resume once the client gate reopens.
-				a.log.WarnContext(ctx, "board webhook: budget exhausted, pausing drain (reconciler heals)",
-					"repo", c.repo, "number", c.number, "head_sha", c.headSHA)
-				return
+				a.hold(order[i:])
+				var wait time.Duration
+				if rle, ok := errors.AsType[*forge.RateLimitError](err); ok {
+					wait = rle.RetryAfter
+				}
+				a.log.WarnContext(ctx, "board webhook: budget exhausted, pausing drain",
+					"repo", c.repo, "number", c.number, "head_sha", c.headSHA, "held", len(a.pending))
+				return wait, true
 			}
 			// Per-event errors log-and-continue (driver.go:96-98 idiom).
 			a.log.WarnContext(ctx, "board webhook: hydrate/sink failed (isolated)",
 				"repo", c.repo, "number", c.number, "head_sha", c.headSHA, "err", err)
 		}
 		if ctx.Err() != nil {
-			return
+			return 0, false
 		}
 	}
+	return 0, false
 }
 
 // resolveAndSink maps a head-SHA CHECKS key to its PR number, then hydrates.
diff --git a/go/internal/ingest/pull_request_sink_test.go b/go/internal/ingest/pull_request_sink_test.go
index a71bf6278..9904ee061 100644
--- a/go/internal/ingest/pull_request_sink_test.go
+++ b/go/internal/ingest/pull_request_sink_test.go
@@ -3,6 +3,7 @@ package ingest
 import (
 	"context"
 	"errors"
+	"runtime"
 	"slices"
 	"sync"
 	"testing"
@@ -446,3 +447,102 @@ func TestReconcileBackfillRetriesAfterRowFailure(t *testing.T) {
 		t.Fatal("repo not marked after the clean retry")
 	}
 }
+
+// TestArmBudgetPauseHoldsPRKeysForRetry: a budget pause keeps the batch's
+// un-hydrated PR keys, including a resolved CHECKS key, and the retry hydrates
+// them once; issue keys are left to the sweep.
+func TestArmBudgetPauseHoldsPRKeysForRetry(t *testing.T) {
+	pulls := &fakePulls{errFor: map[uint64]error{9: &forge.RateLimitError{RetryAfter: 30 * time.Second}}}
+	arm, sink := newPRArm(t, pulls, fakeNumbers{"abc": 12})
+	arm.Enqueue(context.Background(), prEvent(9, changeUpdate))
+	arm.Enqueue(context.Background(), checksEvent())
+	arm.Enqueue(context.Background(), issueEvent("owner/repo", 4, changeUpdate))
+	wait, paused := arm.drainBatch(context.Background(), <-arm.queue)
+	if !paused || wait != 30*time.Second {
+		t.Fatalf("pause = %v wait %v, want paused with the 30s hint", paused, wait)
+	}
+	if len(arm.pending) != 2 {
+		t.Fatalf("held %v, want the two PR keys", arm.pending)
+	}
+
+	pulls.errFor = nil
+	if _, paused := arm.retryPending(context.Background()); paused {
+		t.Fatal("retry paused with the gate open")
+	}
+	got := pulls.readNumbers()
+	slices.Sort(got)
+	if !slices.Equal(got, []uint64{9, 9, 12}) || len(sink.got) != 2 || len(arm.pending) != 0 {
+		t.Fatalf("reads = %v sank %d held %d, want [9 9 12], 2 and 0", got, len(sink.got), len(arm.pending))
+	}
+}
+
+// TestArmHeldKeyNeverPausesFreshBatch: while a PR key is held, a new event for
+// it is left to the timer and a fresh batch hydrates without touching it.
+func TestArmHeldKeyNeverPausesFreshBatch(t *testing.T) {
+	pulls := &fakePulls{errFor: map[uint64]error{9: forge.ErrBudgetExhausted}}
+	arm, sink := newPRArm(t, pulls, nil)
+	arm.Enqueue(context.Background(), prEvent(9, changeUpdate))
+	drainAll(context.Background(), arm)
+	arm.Enqueue(context.Background(), prEvent(9, changeComment))
+	arm.Enqueue(context.Background(), prEvent(10, changeUpdate))
+	drainAll(context.Background(), arm)
+	if got := pulls.readNumbers(); !slices.Equal(got, []uint64{9, 10}) || len(sink.got) != 1 {
+		t.Fatalf("reads = %v sank %d, want [9 10] and 1", got, len(sink.got))
+	}
+	if _, held := arm.pending[boardCoord{repo: "owner/repo", kind: boardKindPR, number: 9}]; !held || len(arm.pending) != 1 {
+		t.Fatalf("held = %v, want only PR 9", arm.pending)
+	}
+}
+
+// TestArmRunRetriesAfterResetHint: Run waits the gate's reset hint, then
+// re-drains the held keys with no new event.
+func TestArmRunRetriesAfterResetHint(t *testing.T) {
+	pulls := &fakePulls{errFor: map[uint64]error{9: &forge.RateLimitError{RetryAfter: 45 * time.Second}}}
+	arm, sink := newPRArm(t, pulls, nil)
+	waits := make(chan time.Duration, 1)
+	fire := make(chan time.Time)
+	arm.after = func(d time.Duration) <-chan time.Time {
+		waits <- d
+		return fire
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan error, 1)
+	go func() { done <- arm.Run(ctx) }()
+
+	defer cancel()
+	arm.Enqueue(ctx, prEvent(9, changeUpdate))
+	select {
+	case d := <-waits:
+		if d != 45*time.Second {
+			t.Fatalf("retry wait = %v, want the 45s hint", d)
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("Run armed no retry after the budget pause")
+	}
+	pulls.mu.Lock()
+	pulls.errFor = nil
+	pulls.mu.Unlock()
+	fire <- time.Time{}
+	waitForReads(t, pulls, 2)
+	cancel()
+	if err := <-done; err != nil {
+		t.Fatalf("Run = %v", err)
+	}
+	if got := pulls.readNumbers(); !slices.Equal(got, []uint64{9, 9}) || len(sink.got) != 1 {
+		t.Fatalf("reads = %v sank %d, want [9 9] and 1", got, len(sink.got))
+	}
+}
+
+// waitForReads waits until pulls has seen n reads; the retry runs on Run's goroutine.
+func waitForReads(t *testing.T, pulls *fakePulls, n int) {
+	t.Helper()
+	deadline := time.After(5 * time.Second)
+	for len(pulls.readNumbers()) < n {
+		select {
+		case <-deadline:
+			t.Fatalf("reads = %v, want %d", pulls.readNumbers(), n)
+		default:
+			runtime.Gosched()
+		}
+	}
+}