diff --git a/.github/workflows/dependency-review.yml b/.github/workflows/dependency-review.yml
index 1b4c1b5..dc44a25 100644
--- a/.github/workflows/dependency-review.yml
+++ b/.github/workflows/dependency-review.yml
@@ -17,7 +17,47 @@ jobs:
       - name: Checkout
         uses: actions/checkout@v7
 
-      - name: Dependency Review
+      # Preferred gate: evaluates dependency changes against GitHub's dependency graph.
+      # Some repositories can have the graph unavailable/disabled. Treat that as a
+      # capability failure, not as permission to skip dependency security validation.
+      - name: GitHub dependency review
+        id: github_dependency_review
+        continue-on-error: true
         uses: actions/dependency-review-action@v4
         with:
           config-file: .github/dependency-review-config.yml
+
+      - name: Setup .NET fallback
+        if: steps.github_dependency_review.outcome == 'failure'
+        uses: actions/setup-dotnet@v6
+        with:
+          dotnet-version: '8.0.x'
+
+      - name: NuGet vulnerability audit fallback
+        if: steps.github_dependency_review.outcome == 'failure'
+        shell: bash
+        run: |
+          set -euo pipefail
+          echo "GitHub dependency review is unavailable or failed; enforcing NuGet audit fallback."
+
+          audit_log="$RUNNER_TEMP/nuget-audit.log"
+          set +e
+          dotnet restore ./PtpLabClock.sln \
+            -p:EnableWindowsTargeting=true \
+            -p:NuGetAudit=true \
+            -p:NuGetAuditMode=all 2>&1 | tee "$audit_log"
+          restore_exit=${PIPESTATUS[0]}
+          set -e
+
+          if [[ "$restore_exit" -ne 0 ]]; then
+            echo "NuGet restore/audit failed with exit code $restore_exit." >&2
+            exit "$restore_exit"
+          fi
+
+          if grep -Eq 'warning NU190[1-4]:' "$audit_log"; then
+            echo "NuGet audit found a package vulnerability (NU1901-NU1904)." >&2
+            grep -E 'warning NU190[1-4]:' "$audit_log" >&2 || true
+            exit 1
+          fi
+
+          echo "NuGet audit completed without NU1901-NU1904 vulnerability warnings."
diff --git a/src/AGENTS.md b/src/AGENTS.md
new file mode 100644
index 0000000..1f4ca3d
--- /dev/null
+++ b/src/AGENTS.md
@@ -0,0 +1,88 @@
+# Source Addendum — PTP Timing, Result Handling, and Diagnostics
+
+The repository root `AGENTS.md` remains authoritative. This scoped addendum applies to `src/**` and strengthens deterministic failure handling and performance rules for protocol, transport, monitor, health, reporting, console, and WPF source code.
+
+## Root-cause workflow
+
+For non-trivial defects use:
+
+REPRODUCE -> TRACE MESSAGE/STATE OWNERSHIP -> ROOT CAUSE -> SMALLEST COHERENT FIX -> REGRESSION TEST -> TIMING/FAILURE CHECK -> BUILD.
+
+If three consecutive patches in the same subsystem still treat symptoms, stop before patch four and re-audit the architecture/state flow. Do not solve packet/state races with arbitrary sleeps, retries, duplicate state, or UI-side correction.
+
+## Exception-free timing-sensitive paths
+
+Expected or recoverable conditions must not use exceptions as routine control flow in:
+- packet serialization/parsing;
+- raw RX/TX loops;
+- Pdelay pairing/response logic;
+- passive monitor processing;
+- sequence/interval/health evaluation;
+- scheduler/timer callbacks;
+- other high-frequency timing paths.
+
+Prefer `TryXxx`, compact typed Result/status records, enums with output values, or nullable values only when failure detail is unnecessary.
+
+Normal conditions such as malformed frame, unsupported message type, sequence discontinuity, missing Follow_Up, timeout, unavailable raw transport, unsuitable adapter, bounded queue saturation, or incomplete peer state must produce explicit status/evidence rather than repeated exception unwinding.
+
+Exceptions from Npcap, OS/network APIs, filesystem/config/reporting, WPF, or third-party infrastructure may still occur. Catch them at meaningful boundaries and convert them into structured application status/diagnostics. Never let infrastructure exceptions unwind through deterministic packet-processing callbacks.
+
+## Defensive packet handling
+
+Treat all captured bytes and configuration fields as untrusted until validated. Before indexing or converting validate:
+- Ethernet/PTP minimum length;
+- EtherType/VLAN layout;
+- messageLength against available bytes;
+- message type/version/domain/flags;
+- timestamp and correction-field bounds;
+- sequence and identity field lengths;
+- integer/range conversions.
+
+Malformed traffic must not terminate monitor or simulator operation. Preserve raw evidence when decoding is uncertain.
+
+## Bounded asynchronous diagnostics
+
+High-rate protocol/timing paths must not perform expensive logging, file writes, report generation, JSON serialization, stack-trace formatting, or synchronous WPF notifications.
+
+Emit only compact structured events/counters such as error code, subsystem, sequence/domain/message type, timestamp/counter, and small numeric context. Transport must be bounded and non-blocking for high-rate producers.
+
+Repeated failures must be aggregated/deduplicated/rate-limited. Queue saturation must use an explicit drop/coalesce policy and retain counters. Diagnostics are observational; their failure must not stall RX/TX, Pdelay response, scheduler operation, passive monitoring, or Demo Mode.
+
+Human-readable messages/report evidence are formatted on a lower-rate/background consumer.
+
+## Timing truth and claims
+
+Software arrival timestamps are evidence about the capture path, not certified PTP accuracy. Do not convert average callback timing into an accuracy claim.
+
+Keep protocol timestamps, software arrival time, scheduler time, and UI presentation time semantically distinct. Any future hardware timestamping must be represented as a separate capability with explicit provenance.
+
+Measure worst-case/sustained behavior where timing-sensitive code changes, not only average latency.
+
+## State ownership
+
+Core owns engine/session/monitor state. Protocol owns wire semantics. Transport owns adapter/raw I/O. Reporting consumes snapshots/evidence and must not become runtime state authority. WPF presents state and issues commands; it must not repair protocol state independently.
+
+Candidate configuration/session state should be validated before activation. Failed reconfiguration must retain last-known-good state where safe or transition to an explicit stopped/faulted state; never leave a half-applied profile.
+
+## Backpressure and UI handoff
+
+Do not render one UI update per received/transmitted packet. Batch/coalesce high-frequency counters and health/status snapshots. Event evidence requiring retention must have an explicit bound/export policy rather than unbounded in-memory growth.
+
+The WPF thread must not perform blocking capture I/O, packet parsing, report generation, PCAP processing, or large serialization.
+
+## Performance/resource evidence
+
+For relevant changes measure:
+- packet/message throughput;
+- scheduler jitter/deadline misses;
+- Pdelay response latency distribution;
+- queue depth/drop/coalesce counters;
+- steady-state allocation rate;
+- CPU and memory growth;
+- UI update latency under sustained monitor load.
+
+Do not add workers, queues, caches, or pools without a demonstrated need and explicit lifecycle owner.
+
+## Definition of done
+
+Protocol/timing changes require deterministic byte-level regression coverage where practical, malformed/failure-path testing, successful repository build/test gates, and PCAP/Wireshark-verifiable evidence when wire behavior changes. Performance-sensitive changes require measured evidence before claiming improvement.