diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 4089db3da..015540029 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -339,6 +339,7 @@ jobs:
         with: { gate: coverage-model }
 
       - name: Run coverage
+        id: run-coverage
         env:
           OUTPUT_ECONOMY_BASE: ${{ github.event_name == 'pull_request' && github.event.pull_request.base.sha || github.event.before }}
         uses: ./.github/actions/run-gate
@@ -355,8 +356,11 @@ jobs:
       # and fails when changed-line coverage < the threshold in
       # scripts/coverage-changed/model.ts. The `coverage-waiver` PR label maps to
       # the waiver env, which skips the failure but still prints the numbers.
+      # Gated on the coverage step's own outcome (#1781 A5): when `Run coverage`
+      # fails, lcov.info is never written, so this step would just re-report that
+      # failure as its own red ("no lcov report") instead of a coverage verdict.
       - name: Enforce changed-line coverage gate
-        if: always() && github.event_name == 'pull_request'
+        if: steps.run-coverage.outcome == 'success' && github.event_name == 'pull_request'
         env:
           AGENT_DEVICE_COVERAGE_WAIVER: ${{ contains(github.event.pull_request.labels.*.name, 'coverage-waiver') }}
         uses: ./.github/actions/run-gate
diff --git a/.github/workflows/perf-nightly.yml b/.github/workflows/perf-nightly.yml
index 00ec20400..81caf11a6 100644
--- a/.github/workflows/perf-nightly.yml
+++ b/.github/workflows/perf-nightly.yml
@@ -1,14 +1,12 @@
 name: Perf Nightly
 
-# End-to-end command perf benchmark (scripts/perf). Scheduled + manual only — perf timing on
-# shared CI runners is noisy, so treat this as a trend/regression signal, not absolute numbers.
-# Reuses the same build artifacts as the device suites: the cached iOS XCUITest runner
-# (setup-apple-runner-build, ios-runner-prebuilt cache) and the Android replay host, and runs the CLI
-# from source via --experimental-strip-types (no dist build), matching the replay workflows.
+# Parked to workflow_dispatch by #1781 A3: it only writes a report and compares nothing, so it
+# structurally cannot catch a regression, and iOS wall-clock medians swing up to +122%
+# night-to-night at n=5 with no consumer reading the report. `pnpm perf` / scripts/perf are
+# unaffected — still runnable on demand. Un-park once a comparison step with multi-night
+# smoothing (someone reading it) exists, or a perf arc needs nightly data.
 
 on:
-  schedule:
-    - cron: '0 4 * * *'
   workflow_dispatch:
     inputs:
       rounds: