diff --git a/FORK.md b/FORK.md
index 827afe10..7fa1ef4b 100644
--- a/FORK.md
+++ b/FORK.md
@@ -75,6 +75,44 @@ Two details worth keeping if this is ever rewritten:
 - `pollTracker` now selects `HNR_HISTORY_POLLS` snapshots instead of 1. Only the HnR check reads the
   extra rows; every other comparison still uses `previousSnapshot`.
 
+## Only credential failures pause a tracker (2.8.9-homelab.7)
+
+Upstream auto-pauses a tracker after `POLL_FAILURE_THRESHOLD` (4) consecutive failures, whatever
+the cause, and a paused tracker only resumes when someone clicks Resume. Two details make that far
+more fragile than it looks:
+
+- A failed poll leaves `lastPolledAt` untouched, so a failing tracker is permanently "overdue" and
+  is retried on **every 5-minute scheduler tick**, not on the hourly poll interval.
+- Four ticks is therefore **20 minutes**. Any outage longer than that pauses the tracker for good.
+
+On 2026-08-16 a home internet outage did exactly that to all six trackers at once. Nothing resumed
+them. The container stayed up the whole time, so the container-level health check stayed green and
+the fault went unnoticed for **33.5 hours** — during which the MyAnonaMouse balance hit its 99,999
+point cap and burned roughly 5,000 points, about 10 GiB of upload credit.
+
+`src/lib/poll-failure-policy.ts` inverts the default. Only a failure a human must actually fix —
+`Authentication failed`, `Session expired`, `Invalid credentials` — can set `pausedAt`. Everything
+else keeps counting failures, so the UI still shows the fault, but is retried forever under
+exponential backoff: 5m, 10m, 20m, 40m, then hourly. Connectivity problems now heal by themselves
+within an hour of the network returning.
+
+Details worth keeping in a rewrite:
+
+- Classification runs on the **output of `sanitizeNetworkError`**, not the raw error, so it matches
+  a small fixed set of phrases rather than guessing at driver-specific text. The unclassified
+  fallback `"Poll failed"` is deliberately transient — that is what the real outage produced, and
+  treating an unknown error as permanent is what caused the incident.
+- Rate-limit and IP-ban errors are transient but jump **straight to the 60-minute cap**, since
+  retrying at the normal cadence is what provokes them.
+- On the transient path `pausedAt` is set to the **column reference** (`paused_at = paused_at`), a
+  no-op self-assign. It must never be a literal, which would clobber a genuine pause.
+- The backoff gate lives in the `pollAllTrackers` overdue filter, keyed off `lastErrorAt`. Without
+  it, removing the pause would mean retrying every 5 minutes forever.
+
+The dashboard side of this incident is fixed separately, in the homelab repo — a heartbeat sender
+that alerts on **snapshot staleness** rather than container liveness, since liveness was never the
+thing worth watching.
+
 ## Workflow edits (the main sync-conflict surface)
 
 `.github/workflows/release.yml` is patched. Upstream's version **cannot publish from a fork** —
diff --git a/package.json b/package.json
index 3dc925e2..f2ca87f2 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "private-tracker-tracker",
-  "version": "2.8.9-homelab.6",
+  "version": "2.8.9-homelab.7",
   "description": "Self-hosted dashboard for monitoring private tracker stats over time",
   "license": "GPL-3.0",
   "repository": {
diff --git a/src/lib/__tests__/circuit-breaker.test.ts b/src/lib/__tests__/circuit-breaker.test.ts
index b8102b4a..32a9332f 100644
--- a/src/lib/__tests__/circuit-breaker.test.ts
+++ b/src/lib/__tests__/circuit-breaker.test.ts
@@ -361,7 +361,36 @@ describe("pollTracker: failure path — consecutiveFailures increment", () => {
     const arg = (failureSetCall as unknown[])[0] as Record<string, unknown>
     // consecutiveFailures must be an actual Drizzle SQL expression, not a literal number
     expect(arg.consecutiveFailures).toBeInstanceOf(SQL)
-    // pausedAt must also be a Drizzle SQL expression (CASE … END), not a raw value
+    // "Connection refused" is transient, so pausedAt must be left exactly as it
+    // is: a self-assign of the column, never a literal that could clobber it.
+    // A transient failure must never be able to pause a tracker, because that
+    // is how a brief outage blinded all six trackers on 2026-08-16.
+    const { trackers } = await import("@/lib/db/schema")
+    expect(arg.pausedAt).toBe(trackers.pausedAt)
+    expect(arg.pausedAt).not.toBeInstanceOf(SQL)
+  })
+
+  it("uses the CASE expression to pause only on a credential failure", async () => {
+    ;(db.select as ReturnType<typeof vi.fn>).mockReturnValue(
+      mockSelectOnce([makeTrackerRow({ consecutiveFailures: 3 })])
+    )
+    // sanitizeNetworkError is mocked to an identity here, so throw the phrase
+    // it would really produce. A dead session is the one class of failure a
+    // human actually has to fix, so it is allowed to pause the tracker.
+    mockFailureAdapter("Session expired")
+    const updateChain = mockUpdateChain([
+      { consecutiveFailures: POLL_FAILURE_THRESHOLD, pausedAt: new Date() },
+    ])
+
+    await pollTracker(1, MOCK_KEY, false)
+
+    const failureSetCall = updateChain.set.mock.calls.find((call: unknown[]) => {
+      const arg = call[0] as Record<string, unknown>
+      return arg.lastError !== undefined
+    })
+    expect(failureSetCall).toBeDefined()
+    const arg = (failureSetCall as unknown[])[0] as Record<string, unknown>
+    expect(arg.lastError).toBe("Session expired")
     expect(arg.pausedAt).toBeInstanceOf(SQL)
   })
 
diff --git a/src/lib/__tests__/poll-failure-policy.test.ts b/src/lib/__tests__/poll-failure-policy.test.ts
new file mode 100644
index 00000000..d836f20e
--- /dev/null
+++ b/src/lib/__tests__/poll-failure-policy.test.ts
@@ -0,0 +1,111 @@
+import { describe, expect, it } from "vitest"
+import {
+  isPermanentPollFailure,
+  isRateLimitFailure,
+  isWithinRetryBackoff,
+  POLL_RETRY_BASE_MS,
+  POLL_RETRY_MAX_MS,
+  pollRetryDelayMs,
+} from "@/lib/poll-failure-policy"
+
+describe("isPermanentPollFailure", () => {
+  it.each(["Authentication failed", "Session expired", "Invalid credentials"])(
+    "treats %s as permanent",
+    (msg) => {
+      expect(isPermanentPollFailure(msg)).toBe(true)
+    }
+  )
+
+  // These are exactly the messages the 2026-08-16 outage produced. If any of
+  // them ever pauses a tracker again, monitoring can blind itself on a blip.
+  it.each([
+    "Poll failed",
+    "Request timed out",
+    "Connection refused",
+    "Host not found",
+    "Host unreachable",
+    "Connection reset",
+    "Proxy connection failed",
+    "API returned 500",
+  ])("treats %s as transient", (msg) => {
+    expect(isPermanentPollFailure(msg)).toBe(false)
+  })
+
+  it("handles null and empty messages", () => {
+    expect(isPermanentPollFailure(null)).toBe(false)
+    expect(isPermanentPollFailure(undefined)).toBe(false)
+    expect(isPermanentPollFailure("")).toBe(false)
+  })
+})
+
+describe("isRateLimitFailure", () => {
+  it("detects tracker-side throttling", () => {
+    expect(isRateLimitFailure("IP temporarily banned by tracker")).toBe(true)
+    expect(isRateLimitFailure("rate limit exceeded")).toBe(true)
+    expect(isRateLimitFailure("Request timed out")).toBe(false)
+  })
+})
+
+describe("pollRetryDelayMs", () => {
+  it("doubles per failure up to the cap", () => {
+    expect(pollRetryDelayMs(1)).toBe(POLL_RETRY_BASE_MS)
+    expect(pollRetryDelayMs(2)).toBe(POLL_RETRY_BASE_MS * 2)
+    expect(pollRetryDelayMs(3)).toBe(POLL_RETRY_BASE_MS * 4)
+    expect(pollRetryDelayMs(4)).toBe(POLL_RETRY_BASE_MS * 8)
+    expect(pollRetryDelayMs(5)).toBe(POLL_RETRY_MAX_MS)
+  })
+
+  it("never exceeds the cap, even for absurd failure counts", () => {
+    expect(pollRetryDelayMs(500)).toBe(POLL_RETRY_MAX_MS)
+    expect(Number.isFinite(pollRetryDelayMs(500))).toBe(true)
+  })
+
+  it("returns zero when there are no failures", () => {
+    expect(pollRetryDelayMs(0)).toBe(0)
+    expect(pollRetryDelayMs(-3)).toBe(0)
+  })
+
+  it("sends rate-limit failures straight to the cap", () => {
+    expect(pollRetryDelayMs(1, "IP temporarily banned by tracker")).toBe(POLL_RETRY_MAX_MS)
+  })
+})
+
+describe("isWithinRetryBackoff", () => {
+  const now = Date.UTC(2026, 7, 17, 12, 0, 0)
+
+  it("is never in backoff with a clean record", () => {
+    expect(
+      isWithinRetryBackoff({ consecutiveFailures: 0, lastErrorAt: new Date(now) }, now)
+    ).toBe(false)
+  })
+
+  it("blocks a retry inside the window and allows it after", () => {
+    const tracker = {
+      consecutiveFailures: 3, // 20 minute backoff
+      lastError: "Request timed out",
+      lastErrorAt: new Date(now - 19 * 60_000),
+    }
+    expect(isWithinRetryBackoff(tracker, now)).toBe(true)
+    expect(isWithinRetryBackoff({ ...tracker, lastErrorAt: new Date(now - 21 * 60_000) }, now)).toBe(
+      false
+    )
+  })
+
+  it("stays due when lastErrorAt is missing or unparseable", () => {
+    expect(isWithinRetryBackoff({ consecutiveFailures: 3, lastErrorAt: null }, now)).toBe(false)
+    expect(isWithinRetryBackoff({ consecutiveFailures: 3, lastErrorAt: "nonsense" }, now)).toBe(
+      false
+    )
+  })
+
+  it("recovers within an hour of a long outage ending", () => {
+    // 40 consecutive failures pins the delay at the 60 minute cap, so the
+    // tracker still gets an attempt every hour and heals on its own.
+    const tracker = {
+      consecutiveFailures: 40,
+      lastError: "Poll failed",
+      lastErrorAt: new Date(now - 61 * 60_000),
+    }
+    expect(isWithinRetryBackoff(tracker, now)).toBe(false)
+  })
+})
diff --git a/src/lib/poll-failure-policy.ts b/src/lib/poll-failure-policy.ts
new file mode 100644
index 00000000..c3bfafc2
--- /dev/null
+++ b/src/lib/poll-failure-policy.ts
@@ -0,0 +1,87 @@
+// src/lib/poll-failure-policy.ts
+//
+// Functions: isPermanentPollFailure, isRateLimitFailure, pollRetryDelayMs
+//
+// Decides whether a failed poll should permanently stop a tracker, or merely
+// slow it down.
+//
+// Background: the circuit breaker used to auto-pause any tracker after
+// POLL_FAILURE_THRESHOLD consecutive failures, regardless of cause, and a
+// paused tracker only resumes when a human clicks Resume. Since a failed poll
+// leaves lastPolledAt untouched, a failing tracker is retried on every 5-minute
+// scheduler tick, so four failures take only ~20 minutes to accumulate. A home
+// internet outage of twenty minutes was therefore enough to permanently
+// disable monitoring for every tracker at once, which is exactly what happened
+// on 2026-08-16: all six trackers paused, nothing resumed them, and the fault
+// went unnoticed for 33.5 hours because the container itself stayed up.
+//
+// The fix inverts the default. Only a failure a human must actually fix -- bad
+// or expired credentials -- pauses a tracker. Everything else is assumed
+// transient and retried forever under exponential backoff, so connectivity
+// problems heal by themselves once the network returns.
+
+/**
+ * Failures that will never resolve on their own, because they need someone to
+ * supply a new credential. Matched against the output of sanitizeNetworkError,
+ * which has already normalised the raw error into a fixed set of phrases.
+ */
+const PERMANENT_FAILURES = [
+  "Authentication failed",
+  "Session expired",
+  "Invalid credentials",
+] as const
+
+/**
+ * Backoff bounds. The base matches the scheduler tick, so the first retry is
+ * unchanged from the old behaviour and only sustained failure slows down. The
+ * cap is the default poll interval -- during a long outage a tracker retries
+ * hourly, which is frequent enough to recover promptly and quiet enough not to
+ * hammer a tracker that may be rate-limiting us.
+ */
+export const POLL_RETRY_BASE_MS = 5 * 60 * 1000
+export const POLL_RETRY_MAX_MS = 60 * 60 * 1000
+
+export function isPermanentPollFailure(message: string | null | undefined): boolean {
+  if (!message) return false
+  return PERMANENT_FAILURES.some((phrase) => message.includes(phrase))
+}
+
+/**
+ * A tracker that is rate-limiting or IP-banning us is transient -- it clears on
+ * its own -- but retrying at the normal cadence is what caused it. These jump
+ * straight to the maximum backoff instead of ramping up to it.
+ */
+export function isRateLimitFailure(message: string | null | undefined): boolean {
+  if (!message) return false
+  return /IP temporarily banned|rate.?limit/i.test(message)
+}
+
+/**
+ * Exponential backoff for a tracker that keeps failing transiently.
+ * 1 failure -> 5m, 2 -> 10m, 3 -> 20m, 4 -> 40m, 5 or more -> 60m.
+ */
+export function pollRetryDelayMs(
+  consecutiveFailures: number,
+  lastError?: string | null
+): number {
+  if (consecutiveFailures <= 0) return 0
+  if (isRateLimitFailure(lastError)) return POLL_RETRY_MAX_MS
+  // clamp the exponent before shifting so a large failure count can't overflow
+  const steps = Math.min(consecutiveFailures - 1, 20)
+  return Math.min(POLL_RETRY_BASE_MS * 2 ** steps, POLL_RETRY_MAX_MS)
+}
+
+/**
+ * True when a tracker is still inside its backoff window and should be skipped
+ * this cycle. A tracker with no recorded failures is always due.
+ */
+export function isWithinRetryBackoff(
+  tracker: { consecutiveFailures: number; lastError?: string | null; lastErrorAt?: Date | string | null },
+  now: number
+): boolean {
+  if (tracker.consecutiveFailures <= 0) return false
+  if (!tracker.lastErrorAt) return false
+  const lastErrorAt = new Date(tracker.lastErrorAt).getTime()
+  if (!Number.isFinite(lastErrorAt)) return false
+  return now - lastErrorAt < pollRetryDelayMs(tracker.consecutiveFailures, tracker.lastError)
+}
diff --git a/src/lib/tracker-scheduler.ts b/src/lib/tracker-scheduler.ts
index 884d37aa..10ba1fef 100644
--- a/src/lib/tracker-scheduler.ts
+++ b/src/lib/tracker-scheduler.ts
@@ -31,6 +31,11 @@ import { localDateStr } from "@/lib/formatters"
 import { POLL_INTERVAL_DEFAULT } from "@/lib/limits"
 import { log } from "@/lib/logger"
 import { dispatchNotifications } from "@/lib/notifications/dispatch"
+import {
+  isPermanentPollFailure,
+  isWithinRetryBackoff,
+  pollRetryDelayMs,
+} from "@/lib/poll-failure-policy"
 import { maskUsername } from "@/lib/privacy"
 import { recordDatabaseSize } from "@/lib/server-data"
 import { HNR_HISTORY_POLLS } from "@/lib/tracker-events"
@@ -452,13 +457,20 @@ export async function pollTracker(
 
     try {
       const now = new Date()
+      // Only a failure a human has to fix may stop polling for good. Transient
+      // failures (timeouts, DNS, refused connections, an unclassified "Poll
+      // failed") keep their failure count for visibility and backoff, but never
+      // set pausedAt -- otherwise a brief outage permanently blinds monitoring.
+      const permanent = isPermanentPollFailure(message)
       const [updated] = await db
         .update(trackers)
         .set({
           lastError: message,
           lastErrorAt: now,
           consecutiveFailures: sql`${trackers.consecutiveFailures} + 1`,
-          pausedAt: sql`CASE WHEN ${trackers.consecutiveFailures} + 1 >= ${POLL_FAILURE_THRESHOLD} THEN ${now.toISOString()}::timestamp ELSE ${trackers.pausedAt} END`,
+          pausedAt: permanent
+            ? sql`CASE WHEN ${trackers.consecutiveFailures} + 1 >= ${POLL_FAILURE_THRESHOLD} THEN ${now.toISOString()}::timestamp ELSE ${trackers.pausedAt} END`
+            : trackers.pausedAt,
           updatedAt: now,
         })
         .where(eq(trackers.id, trackerId))
@@ -479,13 +491,16 @@ export async function pollTracker(
           `Tracker ${trackerId} auto-paused after ${updated.consecutiveFailures} consecutive failures`
         )
       } else if (updated) {
+        const retryInMs = pollRetryDelayMs(updated.consecutiveFailures, message)
         log.info(
           {
             trackerId,
             consecutiveFailures: updated.consecutiveFailures,
             threshold: POLL_FAILURE_THRESHOLD,
+            permanent,
+            retryInMinutes: Math.round(retryInMs / 60_000),
           },
-          `Poll failure ${updated.consecutiveFailures}/${POLL_FAILURE_THRESHOLD} for tracker ${trackerId}`
+          `Poll failure ${updated.consecutiveFailures} for tracker ${trackerId}, retrying in ${Math.round(retryInMs / 60_000)}m`
         )
       }
     } catch (dbError) {
@@ -608,6 +623,24 @@ export async function pollAllTrackers(encryptionKey: Buffer): Promise<void> {
       )
       return false
     }
+    // A failed poll leaves lastPolledAt untouched, so a failing tracker stays
+    // permanently overdue and would otherwise be retried on every 5-minute
+    // tick. Space the retries out instead, so a long outage costs a handful of
+    // attempts per hour rather than twelve.
+    if (isWithinRetryBackoff(tracker, now)) {
+      log.debug(
+        {
+          tracker: tracker.name,
+          consecutiveFailures: tracker.consecutiveFailures,
+          lastError: tracker.lastError,
+          retryInMinutes: Math.round(
+            pollRetryDelayMs(tracker.consecutiveFailures, tracker.lastError) / 60_000
+          ),
+        },
+        "skipping tracker inside retry backoff"
+      )
+      return false
+    }
     const lastPoll = tracker.lastPolledAt?.getTime() ?? 0
     return now - lastPoll >= globalIntervalMs - BATCH_TOLERANCE_MS
   })