diff --git a/.claude/board/EPIPHANIES.md b/.claude/board/EPIPHANIES.md
index 7554dd36..1ee3977f 100644
--- a/.claude/board/EPIPHANIES.md
+++ b/.claude/board/EPIPHANIES.md
@@ -1,3 +1,43 @@
+## E-A-FEATURE-CONDITIONAL-CLASSID-SILENTLY-EMPTIED-A-FIXTURE-1 (2026-08-05, measured)
+
+**A test that hardcodes a classid the domain now selects BY FEATURE does not
+fail loudly — it runs every assertion against an empty projection.** Found by
+finally running `lance-graph-callcenter` under `--features query`, which no
+`-p` sweep and no CI tier had exercised.
+
+**The mechanism, in three links.** (1) `soa_graph::OSINT_GOTHAM.classid` is
+`OSINT_GOTHAM_CLASSID`, which is `NodeGuid::CLASSID_OSINT_V3` (`0x0701_1000`)
+under the default `guid-v3-tail` and only falls back to the V1
+`NodeGuid::CLASSID_OSINT` (`0x0700_0000`) with the feature off. (2)
+`project_snapshot` filters `r.key.classid() == domain.classid` — the codex-P1
+guard that stops a mixed-class board leaking other domains' rows. (3) The
+`graph_table` fixture minted its rows with the V1 constant spelled literally.
+So in a DEFAULT build the filter matched **zero of two rows**, and the test
+compared 0 against 4.
+
+**The second link, which the first one hides.** Fixing only the classid is not
+enough: once the classid is V3, `family_of` dispatches through `family_v2()`,
+so a `NodeGuid::new` key (V1 tail, `family:u24` at bytes 10..13) would carry the
+family in the wrong bits even after the filter passed. The fixture had to move
+to `mint_for(classid_read_mode(c).tail_variant, …)` — which is what the CANON
+already requires of every new mint, and exactly the `ISS-V1-TAIL-RESIDUE` class.
+A fixture that reads BOTH the classid and the tail variant off the domain is
+correct under either feature setting; one that encodes either is a time bomb.
+
+**Why it survived.** Not a DataFusion-54 regression — `num_rows() == 0` is
+decided before DataFusion is involved. It is V1→V3 migration residue that no
+gate ran: the failing pair only surfaces with a feature combination outside the
+default sweep. This is the SECOND instance this session of the same shape
+(the first: `vsa_udfs` / `transcode::ontology_table` returning clippy EXIT=0
+while never compiling, both behind `query` / `query-lite`). The generalization
+worth keeping: **a feature-gated module is not covered by a green run that
+never compiled it, and a feature-CONDITIONAL constant is not pinned by a test
+that spells one of its branches.**
+
+**Fixed:** `crates/lance-graph-callcenter/src/graph_table.rs` — 211/211 green
+under `--features query`. **Not fixed:** nothing yet runs that feature in CI, so
+the coverage hole that hid this is still open.
+
 ## E-LANCE-IS-UPSTREAM-AUTHORITATIVE-1 (2026-08-05, operator-ruled — corrects a P0 in `CLAUDE.md`)
 
 **The lance family is consumed from crates.io upstream and NEVER from a fork —
diff --git a/crates/lance-graph-callcenter/src/graph_table.rs b/crates/lance-graph-callcenter/src/graph_table.rs
index 81285fc5..e17823c9 100644
--- a/crates/lance-graph-callcenter/src/graph_table.rs
+++ b/crates/lance-graph-callcenter/src/graph_table.rs
@@ -115,10 +115,37 @@ pub fn register_graph(
 #[cfg(test)]
 mod tests {
     use super::*;
-    use lance_graph_contract::canonical_node::{EdgeBlock, NodeRow};
+    use lance_graph_contract::canonical_node::{classid_read_mode, EdgeBlock, NodeRow};
     use lance_graph_contract::soa_graph::{project_snapshot, OSINT_GOTHAM};
     use lance_graph_contract::NodeGuid;
 
+    /// Mint an OSINT member key the way the CANON requires: through
+    /// `mint_for(classid_read_mode(c).tail_variant, …)`, with the classid taken
+    /// from the DOMAIN rather than hardcoded.
+    ///
+    /// Both halves matter and both were wrong here before. `OSINT_GOTHAM.classid`
+    /// is `CLASSID_OSINT_V3` under the default `guid-v3-tail` and only falls back
+    /// to the V1 `CLASSID_OSINT` with the feature off; pinning the V1 constant
+    /// made `project_snapshot`'s `classid == domain.classid` filter match ZERO
+    /// rows in a default build, so every assertion below ran against an empty
+    /// snapshot. And once the classid is V3, `family_of` reads the tail through
+    /// `family_v2()` — so a `NodeGuid::new` (V1 tail) key would carry the family
+    /// in the wrong bits even after the filter passed. Reading both the classid
+    /// and the tail variant off the domain keeps the fixture correct under either
+    /// feature setting instead of encoding one of them.
+    fn osint_key(heel: u16, family: u32, identity: u32) -> NodeGuid {
+        NodeGuid::mint_for(
+            classid_read_mode(OSINT_GOTHAM.classid).tail_variant,
+            OSINT_GOTHAM.classid,
+            heel,
+            0,
+            0,
+            0,
+            family,
+            identity,
+        )
+    }
+
     /// Two OSINT members in families 0xA, 0xB; the family-0xA member carries an
     /// out-of-family adapter byte 0x0B → family 0xB. project_snapshot →
     /// GraphSnapshot → arrow tables. End-to-end head → DataFusion.
@@ -127,12 +154,12 @@ mod tests {
         a_edges.out_family[0] = 0x0B;
         let rows = [
             NodeRow {
-                key: NodeGuid::new(NodeGuid::CLASSID_OSINT, 1, 0, 0, 0xA, 1),
+                key: osint_key(1, 0xA, 1),
                 edges: a_edges,
                 value: [0u8; 480],
             },
             NodeRow {
-                key: NodeGuid::new(NodeGuid::CLASSID_OSINT, 2, 0, 0, 0xB, 1),
+                key: osint_key(2, 0xB, 1),
                 edges: EdgeBlock::default(),
                 value: [0u8; 480],
             },