diff --git a/infra/src/lib/stacks/observability-stack.ts b/infra/src/lib/stacks/observability-stack.ts
index c7e20389..635b7c32 100644
--- a/infra/src/lib/stacks/observability-stack.ts
+++ b/infra/src/lib/stacks/observability-stack.ts
@@ -877,7 +877,7 @@ export class ObservabilityStack extends cdk.Stack {
       'UsdStrandedAlarmCount',
       'usd-stranded',
       (count) =>
-        `${count} or more USDT-quoted candles are still at close_usd = 0 more than 48 h after being written, despite a close large enough to price. A zero is indistinguishable from "no data" at ~130 unguarded argMax(close_usd, ...) sites (task 0145), and BE render an empty "--" TVL when nothing priced within 48 h — so this is a value the consumer has already lost, not one that is merely late. MEASURED BASELINE (prod, 2026-08-19): every USDT-quoted hourly candle is priced by 30 h of age — zero unpriced in every age band from 30 h out to 162 h — so this alarm firing past 48 h means real latency has roughly doubled, not that it is marginally slow. USDT is priced through a TWO-HOP chain (usd_rate(USDC) -> the USDT/USDC reference candle's close_usd -> every USDT-quoted candle), so it is structurally one hop behind every other leg; check that chain in order rather than assuming the sweep. A USDT/USDC hourly reference does exist (9-18 buckets/day measured), so its absence is a finding, not the norm. Task 0209 carries the analysis. ⚠️ Do NOT respond by re-running a reset repair — task 0182's own reset CREATED 157 stranded candles by zeroing rows that sat below its pivot reference. Rungs are operator-tunable via config.opsAlarms.usdSanityEscalationCounts.`,
+        `${count} or more USDT-quoted candles are still at close_usd = 0 more than 48 h after being written, despite a close large enough to price. A zero is indistinguishable from "no data" at ~130 unguarded argMax(close_usd, ...) sites (task 0145), and BE render an empty "--" TVL when nothing priced within 48 h, so this is a value the consumer has already LOST, not one that is merely late. KNOWN CAUSE as of 2026-08-20: the USDT pivot has NEVER priced a price_ohlcv_1m row (measured pivot_written = 0 against 1,564,045 peg-written), so this leg has been dark since 2026-08-13 and this alarm stays latched until that is fixed - tasks 0209 (root cause) and 0212 (the peg-valued rows). Verify on _1m, NEVER on a coarse tier: task 0182 repaired the coarse tables directly, so _1h reads clean over a broken _1m. Do NOT re-run a reset repair - 0182 own reset CREATED 157 stranded candles. Rungs tunable via config.opsAlarms.usdSanityEscalationCounts.`,
     );
 
     // Materialized-view drift, on a schedule (task 0204, gap 3). Task 0142 built
@@ -1106,5 +1106,51 @@ export class ObservabilityStack extends cdk.Stack {
     cdk.Tags.of(this).add('Project', 'stellar-prices-api');
     cdk.Tags.of(this).add('ManagedBy', 'cdk');
     cdk.Tags.of(this).add('Environment', config.envName);
+
+    assertAlarmDescriptionsFitCloudWatch(this);
+  }
+}
+
+/**
+ * CloudWatch caps `AlarmDescription` at 1024 characters, and **nothing local
+ * enforces it**: `cdk synth` renders an over-long description happily, the
+ * template is valid CloudFormation, and the request is only rejected by the
+ * CloudWatch API mid-deploy — after some alarms in the stack have already been
+ * created (task 0204, 2026-08-20; three `usd-stranded` rungs at ~1250 chars).
+ *
+ * ⚠️ The alarms in this stack carry deliberately long, runbook-style
+ * descriptions, because an operator reading Slack at 03:00 has nothing else.
+ * That is worth keeping — but it means this ceiling will be hit again, and a
+ * failure discovered at deploy time is the most expensive place to discover it.
+ *
+ * So the check runs at synth: a walk of the construct tree that throws with the
+ * offending alarm and its length. It reads the resolved CloudFormation property
+ * rather than the constructor argument, so descriptions built from tokens or
+ * `Fn::Join` are measured as CloudWatch will actually see them.
+ */
+function assertAlarmDescriptionsFitCloudWatch(scope: Construct): void {
+  const MAX = 1024;
+  const tooLong = scope.node
+    .findAll()
+    .filter((c): c is cloudwatch.Alarm => c instanceof cloudwatch.Alarm)
+    .map((alarm) => {
+      const cfn = alarm.node.defaultChild as cloudwatch.CfnAlarm;
+      const description = cdk.Stack.of(alarm).resolve(
+        cfn.alarmDescription,
+      ) as unknown;
+      const length = typeof description === 'string' ? description.length : 0;
+      return { name: cfn.alarmName, length };
+    })
+    .filter((a) => a.length > MAX);
+
+  if (tooLong.length > 0) {
+    const detail = tooLong
+      .map((a) => `  ${String(a.name)} — ${a.length} chars`)
+      .join('\n');
+    throw new Error(
+      `${tooLong.length} CloudWatch alarm description(s) exceed the ${MAX}-character ` +
+        `API limit and would fail mid-deploy:\n${detail}\n` +
+        'Shorten the description(s); the limit is enforced by CloudWatch, not by synth.',
+    );
   }
 }
diff --git a/lore/1-tasks/active/0204_FEATURE_ops-alarms-missed-an-11h-outage.md b/lore/1-tasks/active/0204_FEATURE_ops-alarms-missed-an-11h-outage.md
index 67b8ef24..1759c6e8 100644
--- a/lore/1-tasks/active/0204_FEATURE_ops-alarms-missed-an-11h-outage.md
+++ b/lore/1-tasks/active/0204_FEATURE_ops-alarms-missed-an-11h-outage.md
@@ -1129,6 +1129,41 @@ supporting measurement — the ~30 h ceiling — was an artefact of 0182's repai
 coverage, not a property of enrichment. Read those sections for the bucket-width
 argument, not for the latency figures.
 
+## 🔴 Deploy attempt 1 FAILED — CloudWatch caps AlarmDescription at 1024 chars
+
+2026-08-20, first `deploy-production-observability`. Three `usd-stranded` rungs
+carried ~1250-character descriptions and CloudWatch rejected them **mid-deploy**,
+after other alarms in the stack had already been created.
+
+⚠️ **Nothing local catches this.** `cdk synth` renders an over-long description
+happily, the template is valid CloudFormation, `cdk diff` showed the expected 12
+additions and 0 deletions, CI was green, and the limit is enforced only by the
+CloudWatch API. **This is AC 4's lesson yet again — "verified by inducing the
+condition, not by reading the CDK" — and this time it bit the deploy itself.**
+
+Two fixes, and the second matters more than the first:
+
+1. **The three descriptions were shortened to ≤ 937 characters.** ⚠️ The text
+   that pushed them over was text **this same day had falsified**: the "MEASURED
+   BASELINE … priced by 30 h of age" claim and the "TWO-HOP chain … check that
+   chain in order rather than assuming the sweep" instruction. Both would have
+   sent a responder down a dead end — the hops were measured healthy while the
+   leg was dark. The rewrite names the real cause (the pivot has never priced a
+   `_1m` row), says the alarm stays latched until [[0209]] is fixed, and warns
+   to verify on `_1m` rather than a coarse tier. **The length limit forced a
+   correction that was owed anyway.**
+2. **`assertAlarmDescriptionsFitCloudWatch()` now runs at synth**, walking the
+   construct tree and throwing with the offending alarm names and lengths. It
+   reads the *resolved* CloudFormation property, so token- or `Fn::Join`-built
+   descriptions are measured as CloudWatch will see them. ⚠️ These alarms carry
+   deliberately long runbook-style descriptions because an operator reading
+   Slack at 03:00 has nothing else — that is worth keeping, which is exactly why
+   the ceiling needs a local guard rather than discipline.
+
+**Verified by inducing it**: padding one description to 2533 chars makes synth
+fail with all three names and lengths; restoring produces a template
+byte-identical to the verified-good one. 34 alarms, max description 937.
+
 ## Future Work
 
 - **Point the peg-applied check at `_1m`** with its own ladder and scan bound —