ONNX Memory: Local embedding backend for cross-session memory

.github/workflows/publish_release_binaries.yml

@@ -143,6 +143,35 @@ jobs:
         --output-dir ./publish
         --version ${{ github.ref_name }}
 
+    # ARM64 cross-compile verification: since ARM64 binaries are built on x64 runners,
+    # we cannot execute them. Instead, verify the build actually produced ARM64 ELF
+    # files (not x64) using the `file` command to detect architecture mismatch. This
+    # catches silent cross-compile failures. See CONTRIBUTING.md § Cross-Platform
+    # Publishing for context.
+    - name: Verify ARM64 binaries are actually ARM64 (not x64)
+      if: matrix.rid == 'linux-arm64'
+      shell: bash
+      run: |
+        set -euo pipefail
+        CLI="./publish/cli/netclaw"
+        DAEMON="./publish/daemon/netclawd"
+
+        for binary in "$CLI" "$DAEMON"; do
+          if [ ! -f "$binary" ]; then
+            echo "ERROR: Expected binary not found: $binary" >&2
+            exit 1
+          fi
+
+          # Check architecture with `file` command
+          file_output=$(file "$binary")
+          if ! echo "$file_output" | grep -q "ARM aarch64"; then
+            echo "ERROR: Binary $binary is not ARM64:" >&2
+            echo "  $file_output" >&2
+            exit 1
+          fi
+          echo "✓ $binary is ARM64"
+        done
+
     - name: Package archives (Unix)
       if: runner.os != 'Windows'
       run: |

.slopwatch/baseline.json

@@ -1,7 +1,7 @@
 {
   "version": 1,
   "createdAt": "2026-05-12T17:20:55.7365203+00:00",
-  "updatedAt": "2026-06-10T19:44:08.3092383+00:00",
+  "updatedAt": "2026-07-08T22:05:54.9570245+00:00",
   "description": "Initial baseline created by 'slopwatch init' on 2026-05-12 17:20:55 UTC",
   "entries": [
     {
@@ -31,24 +31,6 @@
       "message": "Adding warnings to NoWarn: OPENAI001",
       "baselinedAt": "2026-05-12T17:20:55.7420301+00:00"
     },
-    {
-      "hash": "1a29ed65e4ed3efb",
-      "ruleId": "SW004",
-      "filePath": "src/Netclaw.Daemon.Tests/Services/ConfigWatcherServiceTests.cs",
-      "lineNumber": 139,
-      "codeSnippet": "Task.Delay(50, ct)",
-      "message": "Test uses Task.Delay(50) which may indicate a timing-dependent test",
-      "baselinedAt": "2026-05-12T17:20:55.7420338+00:00"
-    },
-    {
-      "hash": "fcb5e461d7f70a7c",
-      "ruleId": "SW004",
-      "filePath": "src/Netclaw.Daemon.Tests/Services/ConfigWatcherServiceTests.cs",
-      "lineNumber": 154,
-      "codeSnippet": "Task.Delay(100, ct)",
-      "message": "Test uses Task.Delay(100) which may indicate a timing-dependent test",
-      "baselinedAt": "2026-05-12T17:20:55.7420454+00:00"
-    },
     {
       "hash": "6ea5c8bbead4b59c",
       "ruleId": "SW004",
@@ -174,6 +156,51 @@
       "codeSnippet": "Fact(SkipUnless = nameof(IsPosix), Skip = \"POSIX-only — matcher routes through BashParser on POSIX\")",
       "message": "Test method 'IsApproved_git_tag_grant_matches_both_version_forms' is disabled: POSIX-only — matcher routes through BashParser on POSIX",
       "baselinedAt": "2026-06-10T19:44:08.3092375+00:00"
+    },
+    {
+      "hash": "687db840a8ff6f35",
+      "ruleId": "SW004",
+      "filePath": "src/Netclaw.Embeddings.Tests/BoundedConcurrencyGateTests.cs",
+      "lineNumber": 32,
+      "codeSnippet": "Task.Delay(20, ct)",
+      "message": "Test uses Task.Delay(20) which may indicate a timing-dependent test",
+      "baselinedAt": "2026-07-08T22:05:54.9567742+00:00"
+    },
+    {
+      "hash": "06c8b73e9ee96bc1",
+      "ruleId": "SW004",
+      "filePath": "src/Netclaw.Embeddings.Tests/BoundedConcurrencyGateTests.cs",
+      "lineNumber": 54,
+      "codeSnippet": "Enumerable.Range(0, 10)\n            .Select(_ => gate.RunAsync(async ct =>\n            {\n                await Task.Delay(5, ct);\n                return Interlocked.Increment(ref completed);\n            }, TestContext.Current.CancellationToken))\n            .ToArray()",
+      "message": "Test uses Task.Delay(?) which may indicate a timing-dependent test",
+      "baselinedAt": "2026-07-08T22:05:54.9568105+00:00"
+    },
+    {
+      "hash": "3d71bddf21a28fee",
+      "ruleId": "SW004",
+      "filePath": "src/Netclaw.Embeddings.Tests/BoundedConcurrencyGateTests.cs",
+      "lineNumber": 57,
+      "codeSnippet": "Task.Delay(5, ct)",
+      "message": "Test uses Task.Delay(5) which may indicate a timing-dependent test",
+      "baselinedAt": "2026-07-08T22:05:54.9568168+00:00"
+    },
+    {
+      "hash": "be802d7249cc2884",
+      "ruleId": "SW004",
+      "filePath": "src/Netclaw.Daemon.Tests/Services/EmbeddingWarmupHostedServiceTests.cs",
+      "lineNumber": 281,
+      "codeSnippet": "Task.Delay(25 * (i + 1))",
+      "message": "Test uses Task.Delay(25 * (i + 1)) which may indicate a timing-dependent test",
+      "baselinedAt": "2026-07-08T22:05:54.9568311+00:00"
+    },
+    {
+      "hash": "233d9d75b339f3f3",
+      "ruleId": "SW004",
+      "filePath": "src/Netclaw.Actors.Tests/Sessions/SQLiteMemoryRecallGateTests.cs",
+      "lineNumber": 456,
+      "codeSnippet": "Task.Delay(Timeout.InfiniteTimeSpan, ct)",
+      "message": "Test uses Task.Delay(Timeout.InfiniteTimeSpan) which may indicate a timing-dependent test",
+      "baselinedAt": "2026-07-08T22:05:54.957+00:00"
     }
   ]
 }

CONTRIBUTING.md

@@ -210,7 +210,7 @@ Prereleases ship to opt-in testers without touching any stable surface.
    Use the **dotted** `beta.N` form (`beta.1`, `beta.2`, … `beta.10`) — never `beta1`.
    A non-dotted identifier compares lexically (so `beta10` would rank below `beta2`), and
    the release version gate rejects it.
-2. Add a `RELEASE_NOTES.md` section for `0.23.0-beta.1`.
+2. Add a `RELEASE_NOTES.md` section for `0.23.0-beta.1` (`## 0.23.0-beta.1 (YYYY-MM-DD)`).
 3. Commit, then tag and push the full version (prefix `-` suffix):
    ```bash
    git tag 0.23.0-beta.1 && git push origin 0.23.0-beta.1

Directory.Packages.props

@@ -116,6 +116,19 @@
   <ItemGroup>
     <PackageVersion Include="BenchmarkDotNet" Version="0.15.8" />
   </ItemGroup>
+  <!-- Memory embeddings (Netclaw.Embeddings + Netclaw.Embeddings.Tests). CPU-only ONNX
+       inference + a pure-managed WordPiece tokenizer, per memory-core-redesign D1: no
+       sidecar process, no native dependency beyond the ONNX Runtime CPU EP that already
+       ships linux-x64 + linux-arm64 native assets. -->
+  <ItemGroup>
+    <PackageVersion Include="Microsoft.ML.OnnxRuntime" Version="1.27.0" />
+    <PackageVersion Include="FastBertTokenizer" Version="1.0.28" />
+    <!-- Pinned to 10.0.10 (matches $(MicrosoftAspNetCoreVersion)), not the lowest version that
+         compiles: Microsoft.Extensions.AI 10.8.0 already pulls System.Numerics.Tensors
+         >= 10.0.10 transitively, and CentralPackageTransitivePinningEnabled requires the
+         central pin to be at least as high as every transitive floor in the graph. -->
+    <PackageVersion Include="System.Numerics.Tensors" Version="10.0.10" />
+  </ItemGroup>
   <!-- Source generators -->
   <ItemGroup>
     <PackageVersion Include="Microsoft.CodeAnalysis.CSharp" Version="5.6.0" />

Netclaw.slnx

@@ -8,6 +8,8 @@
   </Folder>
   <Folder Name="/src/">
     <Project Path="src/Netclaw.Actors/Netclaw.Actors.csproj" />
+    <Project Path="src/Netclaw.Embeddings/Netclaw.Embeddings.csproj" />
+    <Project Path="src/Netclaw.Embeddings.Tests/Netclaw.Embeddings.Tests.csproj" />
     <Project Path="src/Netclaw.Channels/Netclaw.Channels.csproj" />
     <Project Path="src/Netclaw.Channels.Discord/Netclaw.Channels.Discord.csproj" />
     <Project Path="src/Netclaw.Channels.Mattermost/Netclaw.Channels.Mattermost.csproj" />

docs/runbooks/memory-health-and-evals.md

@@ -113,6 +113,90 @@ PY
 
 Then restart the daemon from local binaries before running evals.
 
+## Relevance Gate Health
+
+The relevance gate (`memory-relevance-gate`) is a post-floor cross-encoder
+stage: for each of the (≤3) candidates that already cleared the cosine
+floor, a small ONNX model (`ms-marco-minilm-l-6-v2`) scores `(query,
+candidate)` jointly and drops anything below the calibrated threshold.
+Activation follows `Memory.Embeddings.Enabled` unless
+`Memory.Recall.RelevanceGate.Enabled`/`Threshold` explicitly override it.
+
+1. Run offline diagnostics and review the `Memory Relevance Gate` check:
+
+```bash
+netclaw doctor
+```
+
+   - `PASS` + "disabled (follows Memory.Embeddings.Enabled...)" or "disabled
+     (Memory.Recall.RelevanceGate.Enabled is explicitly false)" — expected,
+     healthy state for any deployment that hasn't opted into embeddings, or
+     that opted out of the gate specifically. Not an error.
+   - `PASS` + "Relevance gate healthy: model '...' provisioned (threshold
+     ...)" — the model is present, hash-verified, and its manifest-carried
+     (or config-overridden) threshold is reported.
+   - `ERROR` + "missing or fails hash verification at `<path>`" — the model
+     was never provisioned or the on-disk artifact doesn't match the pinned
+     SHA-256. Restart the daemon to re-provision if `AutoDownload` is
+     enabled; otherwise provision manually and restart.
+
+2. Check the degradation log line. When the gate is skipped for a turn
+   (model unavailable, its sub-budget exceeded — a 120 ms ceiling clamped to
+   whatever remains of the outer 300 ms `Memory.RecallTimeoutMs` envelope, so
+   a turn where earlier stages already ran long gets less than 120 ms; raised
+   from a fixed 60 ms by a 2026-07 production-canary finding of cold-start
+   timeouts — or recall running in lexical mode because there's no query
+   vector), the coordinator logs a rate-limited marker instead of silently
+   changing what gets injected:
+
+```
+memory_recall_gate_degraded session=<id> reason=<reason> elapsedMs=<ms>
+```
+
+   `reason` is one of `gate_disabled_by_config`, `no_scorer_configured`,
+   `scorer_unavailable`, `sub_budget_exceeded`, or `score_failed:<ExceptionType>`.
+   `elapsedMs` is 0 for the first three (no scoring attempt ever started) and
+   the measured time spent before degrading for the latter two — useful for
+   telling a genuine cold-start/contention timeout apart from an instant
+   failure. Logged at `Warning` when the gate is enabled but a turn still
+   degraded (a genuine runtime condition worth noticing); logged at `Debug`
+   when the gate is off by config (the default, intentional state — not
+   spam). Rate-limited per-reason with the same cooldown as
+   `memory_recall_vector_degraded`, so expect at most one `Warning` line per
+   reason per cooldown window even under sustained degradation, not one per
+   turn.
+
+3. Read `gateScores`/`droppedByGate`/`gateElapsedMs` on `memory_retrieval_final`
+   when diagnosing over- or under-injection or quantifying gate latency
+   margin against the 120 ms ceiling:
+
+```bash
+grep memory_retrieval_final "$HOME/.netclaw/logs/daemon-$(date +%F).log" | tail -20
+```
+
+   - `droppedByGate` — how many of the floor's survivors the gate rejected
+     this turn. `0` on a turn that also injected nothing means the floor
+     itself already filtered everything (or the gate didn't run); a nonzero
+     `droppedByGate` with zero final `injectedCount` means the gate is the
+     reason nothing was injected, not the floor.
+   - `gateScores` — the cross-encoder score for every candidate the gate
+     scored (`id=score`, e.g. `doc-abc123=0.014`), regardless of whether it
+     survived. Compare against the active threshold (config override, or the
+     manifest's calibrated default reported by the doctor check) to see how
+     close a dropped candidate came, or how comfortably a survivor cleared
+     the bar. Absent `gateScores` (empty) on a hybrid-mode turn is itself a
+     signal the gate didn't run for that turn — check for a paired
+     `memory_recall_gate_degraded` line first before assuming a config
+     problem.
+   - Zero `gateScores` and zero `droppedByGate` on a turn is normal whenever
+     the floor itself already produced zero survivors — the gate never runs
+     against an empty candidate set. This is not a gate failure.
+
+See `openspec/changes/memory-relevance-gate/design.md` for the calibration
+procedure (threshold-sweep protocol, model shoot-out, and out-of-sample
+validation numbers) if the operating point ever needs to be re-verified
+against a different relevance model or corpus.
+
 ## Reproducible Memory Score (Non-LLM Judge)
 
 Run the deterministic memory score script:

evals/run-evals.sh

@@ -1137,6 +1137,24 @@ assert_memory_recall_filters() {
     '
 }
 
+# memory-relevance-gate task 2.6: the automated analogue of the shoot-out's "zero-injection
+# accuracy" metric. Off-topic query against the seeded corpus (travel/color/project-alpha/
+# secret-token fixtures) must inject nothing.
+#
+# The eval container does not set Memory.Embeddings.Enabled (default false), so this case
+# exercises the pre-existing lexical-only floor, not the cross-encoder gate itself (that
+# requires an out-of-process download+provisioning step outside this harness's scope — see
+# openspec/changes/memory-relevance-gate/tasks.md task 2.6: "authoring the case is [required];
+# the eval RUN is not required here"). Asserting injectedCount=0 plus the unconditional
+# droppedByGate= field (present on every memory_retrieval_final line regardless of mode --
+# droppedByGate=0 accurately reports "nothing was dropped because nothing reached the gate")
+# keeps this case correct and meaningful in both today's lexical-only default and a future run
+# with embeddings enabled, without needing to touch the eval container's global config.
+assert_memory_relevance_gate_zero_injection() {
+    daemon_log_contains 'memory_retrieval_final.*injectedCount=0' \
+        && daemon_log_contains 'droppedByGate='
+}
+
 # Category 4: Tool Discovery & Use
 assert_tool_discovery() {
     stdout_contains '\[tool:call\] search_tools'
@@ -1761,6 +1779,9 @@ run_all() {
     run_case memory_recall_filters "candidate selection with score filtering" \
         "Tell me about my travel preferences"
 
+    run_case memory_relevance_gate_zero_injection "off-topic query injects nothing, gate marker logged" \
+        "What is the boiling point of tungsten in degrees Celsius?"
+
     end_category
 
     # ── Category 4: Tool Discovery & Use ──