feat(adapters): RlmAdapter + Qwen-7B/Llama-70B baselines

docs/ideas.md

@@ -430,6 +430,44 @@ shape future adapter work and category implementations:
   and `baselines/jp_realm_v0_1_*.json` made it obvious where to
   add `jp_realm_v0_2`, `family_palace_v0_1`, etc. Don't refactor.
 
+## RlmAdapter — research scaffold (2026-04-26)
+
+`sme/adapters/rlm_adapter.py` ships an adapter that treats RLM
+(jphein/rlm fork of alexzhang13/rlm) as the read-side orchestrator:
+the LLM itself decides when to call `mempalace_search`, with what
+queries, and how to compose results. familiar's deterministic
+retrieve→rerank→decay→compress pipeline becomes the *baseline* this
+adapter is benchmarked against.
+
+**Design:** RLM gets `mempalace_search` registered as a
+`custom_tools` callable. The adapter wraps that callable to capture
+every search result into a per-query buffer. After `rlm.completion()`
+returns, the buffer's contents become `context_string` (in tool-call
+order) and `retrieved_entities` (one Entity per drawer). The Cat 1 /
+retrieve substring scorer measures whether `expected_sources` ended
+up in `context_string` — same contract as every other adapter.
+
+**Test coverage** (`tests/test_rlm_adapter.py`, 5 tests): tool-call
+aggregation, capture-buffer reset across queries, error-dict
+graceful handling on palace network failure, empty graph snapshot
+(Cat 8 N/A for RLM), ingest_corpus skipped.
+
+**To benchmark live:**
+```bash
+PORTKEY_API_KEY=... PALACE_DAEMON_URL=http://disks.jphe.in:8085 \
+PALACE_API_KEY=... \
+venv/bin/sme-eval retrieve --adapter rlm \
+    --questions sme/corpora/jp_realm_v0_1/questions.yaml \
+    --json baselines/jp_realm_v0_1_rlm_$(date +%Y%m%d).json
+```
+
+A live benchmark will reveal whether RLM-orchestration recovers
+recall on questions familiar misses, plateaus at the same number,
+or regresses — the answer determines whether v0.4 should
+productionize RLM into familiar's chat path. Without that data,
+the design spec's "RLM and familiar are complementary" hypothesis
+stays a hypothesis.
+
 ## What's next
 
 ### Categories that aren't implemented yet

docs/sme_spec_v8.md

@@ -176,9 +176,16 @@ class SMEAdapter(ABC):
         Returns: {'type': 'declared'|'readme'|'inferred',
                   'schema': [...], 'documentation': str}"""
         return {'type': 'inferred', 'schema': [], 'documentation': ''}
+
+    def get_harness_manifest(self) -> list[HarnessDescriptor]:
+        """Return invocation surfaces this memory system exposes.
+        Used by Category 9 (Harness Integration). Adapters that don't
+        expose a harness surface (pure library APIs) return [] — Cat 9b
+        reports empty_manifest=True rather than crashing."""
+        return []
 ```
 
-Three required methods. That's the minimum viable adapter. `get_flat_retrieval` and `get_ontology_source` have defaults — SME fills in its own flat baseline and infers ontology from the graph if the adapter doesn't provide them.
+Three required methods. That's the minimum viable adapter. `get_flat_retrieval`, `get_ontology_source`, and `get_harness_manifest` have defaults — SME fills in its own flat baseline, infers ontology from the graph, and treats an empty harness manifest as "Cat 9 not applicable" when the adapter doesn't provide them.
 
 ### Default Adapters
 

pyproject.toml

@@ -47,6 +47,17 @@ ladybugdb = [
 neo4j = [
     "neo4j>=5.0",
 ]
+# RlmAdapter — Recursive Language Models. Pinned to jphein/rlm (fork
+# of alexzhang13/rlm) because the baselines/jp_realm_v0_1_rlm_*.json
+# artifacts were produced against this fork; reproducing them on a
+# fresh clone needs the same code path, even when the package surface
+# is currently identical to upstream. Distribution name is `rlms`;
+# importable module is `rlm`. Requires Python >=3.11 (one minor
+# higher than SME's >=3.10 floor), so this extra is opt-in and not
+# part of `all`. Tests are guarded with `pytest.importorskip`.
+rlm = [
+    "rlms @ git+https://github.com/jphein/rlm.git",
+]
 all = [
     "sme-eval[topology,viz,ladybugdb,neo4j]",
 ]

sme/adapters/base.py

@@ -4,7 +4,7 @@
 never touches a database directly — it talks to this thin interface.
 
 Three required methods: ingest_corpus, query, get_graph_snapshot.
-Two optional: get_flat_retrieval, get_ontology_source.
+Three optional: get_flat_retrieval, get_ontology_source, get_harness_manifest.
 """
 
 from __future__ import annotations
@@ -116,7 +116,7 @@ class QueryResult:
 class SMEAdapter(ABC):
     """Implement this for your database/memory system.
 
-    Three required methods. Two optional.
+    Three required methods. Three optional.
     """
 
     # --- Required ------------------------------------------------------