feat+revert(v0.22): domain enumeration in graph context; strategy text reverted

PHASE 2.1 (kept): build_graph_context now emits a "Domains in this
corpus: krra (90125), x2bee (19843), assort (13909)" line whenever
the corpus has 2+ distinct properties._domain_id values. Direct SQL
DISTINCT (not list_nodes sampling) so the count is accurate even
when nodes are stored in domain-contiguous order.

PHASE 2.2 (reverted strategy text, kept enumeration): tested a
3-step strategy prompt block ("identify domains → fan-out → verify").
Re-bench cross-domain shows:

  v0.22 hits 3/12 (same as v0.21), but degraded partial coverage:
    xd001: miss → hit  (assort=6 unlocked) ✓
    xd008: hit → miss  (assort=10 → 0)    ✗
    xd010: 2/3 → 1/3 partial coverage     ✗
    xd012: krra=83 → found=0 catastrophic ✗

Net zero hits, net negative on partial coverage — same brittle
deterministic-prompt-shift dynamic seen in v0.20. Strategy block
reverted; only the factual enumeration line remains (zero behavioural
push, agent CAN read it but isn't being steered).

Pattern (3 iterations now): adding text to AGENT_SYSTEM at temp=0/
seed=42 is a coin-flip per query. v0.19 +4, v0.20 -5 (reverted as
v0.20.1 +1), v0.22 +0. Conclusion: prompt-tuning won't move
cross-domain coverage. Need tool-level fan-out (Phase 2.3): modify
search/deep_search to split into per-domain sub-searches internally
so the agent makes one call and gets multi-domain results without
behaviour change.

Tests: 983 pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: SonAIengine

CHANGELOG.md

@@ -6,6 +6,44 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+### Measured — v0.22 Phase 2.1/2.2: domain-aware prompt is NET ZERO (reverted strategy text)
+
+Tried lifting cross_domain coverage by adding multi-domain awareness
+to the agent system prompt — both the factual enumeration of
+distinct ``_domain_id`` values AND a 3-step strategy block telling
+the agent to identify domains, fan out parallel searches, verify
+coverage. Same dynamic seen in v0.20: the prompt addition rerouted
+deterministic decoding paths and broke as many queries as it helped.
+
+| qid | v0.21 baseline | **v0.22 with strategy** | delta |
+|---|---|---|---|
+| xd001 | miss (krra=112, assort=0) | **hit** (krra=136, assort=6) | flipped to hit ✓ |
+| xd008 | **hit** (krra=100, assort=10) | miss (krra=48, assort=0) | flipped to miss ✗ |
+| xd010 | partial (krra+x2bee 2/3) | worse (krra only 1/3) | regression |
+| xd012 | partial (krra=83) | **found=0** | catastrophic |
+| **total hits** | **3 / 12** | **3 / 12** | **0** |
+
+Same hit count, worse partial-coverage signal on 3-domain queries.
+This is now the third iteration where adding text to the agent
+system prompt at temp=0/seed=42 has produced net-neutral or net-
+negative results (v0.20 cursor follow-through was −5; v0.22 domain
+strategy is −0 hits / −2 partial).
+
+**Conclusion**: agent prompt tuning is fundamentally unreliable at
+deterministic sampling. Each prompt change is a coin-flip per query.
+The factual domain enumeration is preserved (zero behavioural risk —
+the agent CAN see ``Domains in this corpus: krra (90125), x2bee
+(19843), assort (13909)`` in the graph metadata block) but the 3-step
+strategy block is reverted.
+
+**Phase 2.3 hypothesis** (next): tool-level fan-out instead of prompt
+instruction. Modify ``search``/``deep_search`` to detect a multi-
+domain corpus and internally split into per-domain sub-searches.
+Agent makes one call, gets multi-domain results without any
+behavioural change. The per-domain sub-search bypasses the FTS
+ranking bias that lets a single dominant category (KRRA's ``ESG 및
+지속가능성``) crowd out content from other domains.
+
 ### Measured — v0.21 Phase 1.5/1.6: cross-domain federation bench (3/12 = 25 % baseline)
 
 End-to-end demo of the Phase 1 stack — Phase 1.4 MetaCorpus combiner +

src/synaptic/search_session.py

@@ -236,6 +236,34 @@ async def build_graph_context(backend: StorageBackend) -> str:
         # Total counts
         total_docs = await backend.count_nodes(kind=None)
 
+        # Per-domain breakdown (Phase 2.1) — surface distinct
+        # ``properties._domain_id`` values so the agent can plan
+        # cross-domain queries. Empty when the corpus has only one
+        # (or no) domain tag — back-compat with single-domain corpora.
+        domains_summary = ""
+        try:
+            # Direct SQL avoids the list_nodes(limit=10K) sampling bias
+            # that returns one domain's worth of contiguous rows on a
+            # MetaCorpus where domains are ordered. Falls back silently
+            # for non-sqlite backends.
+            db_method = getattr(backend, "_db", None)
+            domain_counts: dict[str, int] = {}
+            if callable(db_method):
+                db = db_method()
+                cur = await db.execute(
+                    "SELECT json_extract(properties_json, '$._domain_id') AS dom, COUNT(*) "
+                    "FROM syn_nodes WHERE dom IS NOT NULL GROUP BY dom ORDER BY 2 DESC LIMIT 10"
+                )
+                for dom, cnt in await cur.fetchall():
+                    if dom:
+                        domain_counts[dom] = int(cnt)
+                await cur.close()
+            if len(domain_counts) >= 2:
+                parts = [f"{d} ({c})" for d, c in domain_counts.items()]
+                domains_summary = ", ".join(parts)
+        except Exception:
+            pass
+
         # Count nodes by kind to distinguish document vs structured graphs.
         # Structured entities are identified by the ``_table_name`` property
         # stamped by TableIngester / DbIngester — raw ENTITY nodes from
@@ -256,6 +284,16 @@ async def build_graph_context(backend: StorageBackend) -> str:
             "Use category names above as the 'category' parameter in search.",
         ]
 
+        # Multi-domain corpus → just enumerate the domains. Verbose
+        # strategy text was tested in v0.22 and net-zero on hit-rate
+        # while degrading partial coverage on 3-domain queries — same
+        # deterministic-prompt-shift dynamic seen at v0.20. Keeping
+        # only the factual enumeration so the agent CAN see them
+        # without being pushed toward a specific decoding path.
+        if domains_summary:
+            lines.append("")
+            lines.append(f"Domains in this corpus: {domains_summary}")
+
         # --- Structured data: table schemas ---
         # Detect tables from _table_name property and sample columns.
         structured_row_count = 0