diff --git a/.context/experiments/bitmap-pushdown.md b/.context/experiments/bitmap-pushdown.md
index 67f88d8..2dac578 100644
--- a/.context/experiments/bitmap-pushdown.md
+++ b/.context/experiments/bitmap-pushdown.md
@@ -1,6 +1,15 @@
 # Experiment 1.5 — Extending DataFusion dynamic-filter-pushdown to bitmap shape (code-dive)
 
-**Ticket:** MR-925 §1.5 (validates MR-737 §5.6, §5.7 / Open Q3).
+**Ticket:** MR-925 §1.5 (validates MR-737 §5.3 "Sideways Information Passing
+(SIP) — extends DataFusion's dynamic filter pushdown" + §5.6 "Capability-bearing
+storage trait", with implications for §5.7 cost-model).
+
+**§-numbering note:** the original MR-925 cross-reference said "§5.3 / §5.10".
+On a full re-read of MR-737: §5.3 is the SIP design and is the correct primary
+§ for this experiment; §5.10 in current MR-737 is "First-class scores and rank
+fusion" (unrelated). The capability seam that the bitmap pushdown traverses
+on the storage side is §5.6 (`scan_by_key_set` capability), so this writeup
+also produces deltas for §5.6.
 **Type:** Code-dive only (no prototype crate).
 **Substrate pin:** DataFusion 52.5.
 **Date:** 2026-05-12.
diff --git a/.context/experiments/custom-lance-index.md b/.context/experiments/custom-lance-index.md
index 16c23e4..125b5e0 100644
--- a/.context/experiments/custom-lance-index.md
+++ b/.context/experiments/custom-lance-index.md
@@ -1,6 +1,14 @@
 # Experiment 1.2 — Custom Lance index plugin from outside the lance crate
 
-**Ticket:** MR-925 §1.2 (validates MR-737 §5.4, §5.5).
+**Ticket:** MR-925 §1.2 (validates MR-737 §5.4 "Persisted CSR adjacency as
+Lance index plugin" + §5.16 "Index rebuild orchestration — stateless reconciler").
+
+**§-numbering note:** the original MR-925 cross-reference said "§5.4, §5.5".
+On a full re-read of MR-737: §5.4 is correct (this is the plugin section);
+§5.5 is "Stable row IDs as graph IDs" — *not* the reconciler. The reconciler
+is §5.16 ("Index rebuild orchestration"). Section references in the body have
+been corrected; §5.5 is *indirectly* in scope only because the experiment
+leans on stable row IDs.
 **Prototype:** `validation-prototypes/custom-lance-index/` (long-lived branch).
 **Substrate pin:** Lance 4.0.1 (matched by cargo to 4.0.0 spec). Lance 4.0.1 internally pulls roaring 0.11 and prost-types 0.14; the workspace deps were lifted to match.
 **Date:** 2026-05-12.
@@ -128,8 +136,8 @@ is closed to us). Any custom-index reconciler we ship has to:
   - emit `Operation::CreateIndex { new_indices: vec![updated], removed_indices: vec![old] }`
     to re-publish the index with the updated bitmap.
 
-This is *consistent with* the §5.5 reconciler pattern in MR-737, so it's
-not a blocker — but the writeup of §5.5 should explicitly say "the
+This is *consistent with* the §5.16 reconciler pattern in MR-737, so it's
+not a blocker — but the writeup of §5.16 should explicitly say "the
 reconciler also owns fragment coverage diffs, not just file content".
 
 ### F5. Compaction does not move our index. ⚠️
@@ -203,7 +211,7 @@ paths forward:
 both. The current MR-737 wording implies path (1) is available; this
 experiment proves it is not.
 
-§5.5 (reconciler pattern) is unaffected by this finding — but it must
+§5.16 (reconciler pattern) is unaffected by this finding — but it must
 expand to explicitly own `fragment_bitmap` recomputation across all
 mutating operations, since with path (2) or path (3) we are the only
 party that knows the index's row coverage.
diff --git a/.context/experiments/custom-operator.md b/.context/experiments/custom-operator.md
index 8bd5f43..5e53c8c 100644
--- a/.context/experiments/custom-operator.md
+++ b/.context/experiments/custom-operator.md
@@ -1,6 +1,16 @@
 # Experiment 1.3 — Custom UserDefinedLogicalNode + ExecutionPlan e2e
 
-**Ticket:** MR-925 §1.3 (validates MR-737 §5.3, §5.10).
+**Ticket:** MR-925 §1.3 (validates MR-737 §5.1 "Unified IR", §5.2 "Factorized IR",
+§5.11 "Substrate choice — DataFusion vs. custom executor (A)", and §5.12
+"Mutation IR" — all of which rely on `UserDefinedLogicalNode` + `ExecutionPlan`
+surviving the optimizer end-to-end).
+
+**§-numbering note:** the original §1.3 cross-reference in MR-925 cited "§5.3,
+§5.10". On a full re-read of MR-737: §5.10 is "First-class scores and rank
+fusion (open-world over modalities)" (not "operators survive the optimizer"),
+and the custom-operator e2e contract is actually shared across §5.1, §5.2,
+§5.11, and §5.12. §5.3 (SIP) is the *first* operator that consumes the contract
+— valid — but the contract itself is broader than §5.3 alone.
 **Prototype:** `validation-prototypes/custom-operator/`.
 **Substrate pin:** DataFusion 52.5 (matched to omnigraph workspace).
 **Date:** 2026-05-12.
@@ -168,7 +178,7 @@ These are not blockers but should be noted for the §11 RFC-body delta:
    the RFC that graph operators must explicitly choose their output
    partitioning rather than inheriting.
 
-## Decision impact on MR-737 §5.3 and §5.10
+## Decision impact on MR-737 §5.1, §5.2, §5.3, §5.11, §5.12
 
 **§5.3 is achievable on DataFusion 52.5 as written.** The
 `UserDefinedLogicalNode`/`ExecutionPlan` surface is fully sufficient
@@ -183,9 +193,9 @@ NeighborSetIntersect, etc.). The only edits needed in §5.3:
     operators implementing it accurately, not punting to
     `Statistics::new_unknown`.
 
-**§5.10 ("operators survive the optimizer + execute correctly")**:
-The composition test (E4) plus the metrics test (E5) cover this. No
-deltas needed.
+**§5.1 / §5.2 / §5.11 / §5.12 ("custom operators survive the optimizer +
+execute correctly")**: The composition test (E4) plus the metrics test (E5)
+cover this. No deltas needed.
 
 ## Caveats
 
diff --git a/.context/experiments/sip-format-bench.md b/.context/experiments/sip-format-bench.md
index fd7dd2c..d8f76e9 100644
--- a/.context/experiments/sip-format-bench.md
+++ b/.context/experiments/sip-format-bench.md
@@ -1,6 +1,15 @@
 # Experiment 1.4 — Roaring bitmap variant for u64 row IDs (SIP wire format)
 
-**Ticket:** MR-925 §1.4 (validates MR-737 §5.6, §5.8 / Open Q4).
+**Ticket:** MR-925 §1.4 (validates MR-737 §5.3 "Sideways Information Passing
+(SIP)" wire format + §10 Open Q3 "SIP wire format").
+
+**§-numbering note:** the original MR-925 cross-reference said "§5.6, §5.8 /
+Open Q4". On a full re-read of MR-737: §5.3 is SIP (the wire format is named
+in Open Q3); §5.6 is the storage trait that *consumes* SIP via `sip_mask` /
+`key_set` on `ScanRequest`; §5.8 is "Tiering via Lance base paths" (unrelated
+to SIP). Open Q4 is "CSR index format inside Lance" (also unrelated). The
+correct primary mapping is **§5.3 + §10 Open Q3**, with downstream
+implications for §5.6 (the capability advertisement).
 **Prototype:** `validation-prototypes/sip-format-bench/`.
 **Substrate pin:** `roaring = "0.11"` (matched to lance-table dependency).
 **Date:** 2026-05-12.
@@ -10,7 +19,7 @@
 ## Hypothesis
 
 For propagating row-ID side-information predicates (SIPs) between operators —
-the §5.6 dynamic-filter-pushdown wire format — Roaring bitmaps over u64
+the §5.3 dynamic-filter-pushdown wire format — Roaring bitmaps over u64
 (`RoaringTreemap`) are the right encoding when row IDs cluster by Lance
 fragment (which they do). For random u64s, Roaring is *not* the right
 choice.
@@ -148,7 +157,7 @@ At `n=1M` dense, encoding takes **37ms**, decoding takes **0.02ms**.
 For "build once, read many" wire-format use, this is fine. But if the
 SIP is built mid-pipeline (e.g. from a `FilterExec`'s output IDs) and
 intersected immediately with another payload, the build cost dominates.
-The §5.6 RFC should clarify: SIPs are produced at *probe-build time* on
+The §5.3 RFC should clarify: SIPs are produced at *probe-build time* on
 the hash-join build side, where 37ms is amortized across the entire
 probe phase.
 
@@ -176,9 +185,9 @@ construction.
 
 Default fallback (for non-row-ID u64s): **varint-delta**.
 
-## Decision impact on MR-737 §5.6 and §5.8
+## Decision impact on MR-737 §5.3 + §10 Open Q3 (and downstream §5.6)
 
-**§5.6 (SIP wire format) — concrete choice:**
+**§5.3 (SIP wire format) — concrete choice:**
 
 > ROW_ID_SIP wire format := length-prefixed roaring `serialize_into` bytes
 > with a 1-byte format-tag prefix. Tag values: `0x01` = Roaring (u64
@@ -189,7 +198,7 @@ Default fallback (for non-row-ID u64s): **varint-delta**.
 This makes the wire format extensible while picking a default that
 matches the dominant workload.
 
-**§5.8 / Open Q4 — answered:**
+**Open Q3 — answered:**
 
 The RFC's Q4 ("can we share the SIP filter between operator stages by
 serializing roaring bytes?") is **yes for row-ID payloads**.
diff --git a/.context/experiments/stable-row-id-compaction.md b/.context/experiments/stable-row-id-compaction.md
index e1ffa9d..58016a0 100644
--- a/.context/experiments/stable-row-id-compaction.md
+++ b/.context/experiments/stable-row-id-compaction.md
@@ -1,6 +1,17 @@
 # Experiment 1.7 — Stable-row-id-aware indices survive compaction (code-dive + small repro plan)
 
-**Ticket:** MR-925 §1.7 (validates MR-737 §5.4, §5.10 / Open Q6).
+**Ticket:** MR-925 §1.7 (validates MR-737 §5.4 "Persisted CSR adjacency as Lance
+index plugin" + §5.5 "Stable row IDs as graph IDs").
+
+**§-numbering note (added on re-read of MR-737):** MR-925's original §1.7 cross-
+reference cited "§5.8 / Open Q7" / "§5.10". On a full read of MR-737, §5.10 is
+"First-class scores and rank fusion" (NOT custom index types), §5.4 is "Persisted
+CSR adjacency as Lance index plugin" (which contains the custom-index-type seam),
+and §5.5 is "Stable row IDs as graph IDs" (which flags the experimental status of
+"Stable Row ID for Index" in lance-4.0.x). The corrected mapping for §1.7 is
+**§5.4 + §5.5** and the MR-737 §5.5 substrate caveat that "`Stable Row ID for
+Index` is documented as experimental in lance-4.0.x" is the immediate caveat for
+this experiment.
 **Type:** Code-dive plus a planned small repro (not yet built; specified for Phase 0 entry).
 **Substrate pin:** Lance 4.0.1, lance-index 4.0.1.
 **Date:** 2026-05-12.
@@ -9,10 +20,13 @@
 
 ## Question
 
-MR-737 §5.4 (graph topology index) and §5.10 (custom index types for graph
-adjacency) both depend on the assumption that a custom index — i.e. our
-own CSR/CSC adjacency lists keyed by source-table row IDs — **continues
-to point at the right rows after the source table is compacted.**
+MR-737 §5.4 ("Persisted CSR adjacency as Lance index plugin") and §5.5 ("Stable
+row IDs as graph IDs") both depend on the assumption that a custom CSR/CSC
+adjacency index keyed by source-table row IDs **continues to point at the right
+rows after the source table is compacted.** §5.5 explicitly flags the substrate
+caveat: "Stable Row ID for Index" is **experimental** in lance-4.0.x; confirming
+whether our created indices opt into stable-row-id mode is a follow-up worth
+doing before MR-848 (index reconciler) lands.
 
 Lance's compaction (`compact_files`) consolidates fragments, which on the
 non-stable row-ID scheme renumbers row addresses. The question:
@@ -253,19 +267,22 @@ demonstrates end-to-end survival. Specification:
 Estimated effort: 1–2 days. **Defer to Phase 0**; the code-dive
 already justifies §5.4 as feasible without the repro.
 
-## Decision impact on MR-737 §5.4 and §5.10
+## Decision impact on MR-737 §5.4 and §5.5
 
-**§5.4 (graph topology index) is feasible on Lance 4.0.1 with stable
-row IDs (Path B):**
+**§5.4 (persisted CSR adjacency as Lance index plugin) is feasible on Lance
+4.0.1 with stable row IDs (Path B):**
 
 - No Lance plugin-registry dependency; we drive remapping ourselves.
 - The custom topology dataset stores stable row IDs end-to-end; the
   bulk of compaction-induced changes don't require remap.
 - Path A (Lance-managed remapping) is a follow-up improvement
-  contingent on the §1.2 contribution.
+  contingent on the §1.2 plugin-registry contribution.
 
-**§5.10 (custom index types):** No new findings beyond §1.2. The
-`ScalarIndex::remap` contract is sufficient and stable.
+**§5.5 (stable row IDs as graph IDs):** The MR-737 substrate caveat
+("`Stable Row ID for Index` is experimental in lance-4.0.x") still
+stands. The small repro in §5 above is the way to confirm opt-in;
+until it runs, treat §5.5 as substrate-positive but not yet validated
+for index-side stable IDs.
 
 **Open Q6 ("survive compaction"):** Answered yes. The recommendation is
 **Path B for v1, Path A for v2**. RFC §5.4 should specify the
diff --git a/.context/experiments/txn-branches-cost.md b/.context/experiments/txn-branches-cost.md
index dbb92fa..cfe001e 100644
--- a/.context/experiments/txn-branches-cost.md
+++ b/.context/experiments/txn-branches-cost.md
@@ -1,6 +1,25 @@
 # Experiment 1.6 — Lance native per-table txn branches (code-dive, cost model)
 
-**Ticket:** MR-925 §1.6 (validates MR-737 §5.11, §5.12 / Open Q5).
+**Ticket:** MR-925 §1.6 (validates MR-737 §5.12 "Mutation IR, write planner,
+and external sources").
+
+**§-numbering note (added on re-read of MR-737):** MR-925's original §1.6 cross-
+reference cited "§5.11, §5.12 / Open Q5". On a full read of MR-737:
+
+- **§5.11** is "Substrate choice — DataFusion vs. custom executor — RECOMMENDED
+  (A)" — resolved 2026-05-11, not about per-table branches.
+- **§5.12** "Mutation IR, write planner, and external sources" is where the
+  *implementation note* says "writes go to per-table Lance native branches
+  (storage-layer staging invisible above the TableProvider/Dataset boundary),
+  success → fast-forward, failure → drop branches. NOT `restore()`-based rollback"
+  (also called out in §0.5 item 9). This is what §1.6 actually validates.
+- **Open Q5** in the current MR-737 §10 is "What 'extension rate' means under
+  filters" — unrelated. The MR-925 cross-reference is to a pre-2026-05-11
+  numbering where Q5 covered per-table branches; that question is now folded
+  into §5.12 and resolved per Ragnor 2026-04-29.
+
+Corrected mapping for §1.6: **§5.12 (mutation-IR per-table branches)** — cost
+model + recommendation.
 **Type:** Code-dive only — no prototype crate.
 **Substrate pin:** Lance 4.0.1.
 **Date:** 2026-05-12.
@@ -204,11 +223,12 @@ cheaper.
    *branches*, not just zombie *commits*. This is more state to track,
    not less.
 
-## Decision impact on MR-737 §5.11 and §5.12
+## Decision impact on MR-737 §5.12
 
-**§5.11 (per-table txn isolation):** Lance-native branches **can**
-implement this, but the steady-state cost is essentially the same as
-the current lazy-graph-branch model, and the abort-path cost is *higher*.
+**§5.12 (mutation IR with per-table branches):** Lance-native branches **can**
+implement the per-table staging shape that §5.12 prescribes, but the steady-state
+cost is essentially the same as the current lazy-graph-branch model, and the
+abort-path cost is *higher*.
 The conceptual clarity argument is real but not load-bearing — both
 models provide the same isolation guarantee.
 
@@ -236,7 +256,7 @@ Lance-native per-table branches only if:
     "lazy" so most branches don't fork most tables).
 
 For Phase 0, the deliverable is a **clear specification** of which model
-MR-737 §5.11/§5.12 prescribes. The cost analysis above suggests
+MR-737 §5.12 prescribes. The cost analysis above suggests
 specifying the current model.
 
 ## Caveats and follow-ups
@@ -255,4 +275,4 @@ specifying the current model.
 - **Forbidden APIs file** (`crates/omnigraph/tests/forbidden_apis.rs:57`)
   excludes `.delete_branch(` from the over-match check — there's
   intent in the codebase to allow these calls. Worth a follow-up read
-  to see if MR-737 §5.11 has already opened the door.
+  to see if MR-737 §5.12 has already opened the door.
diff --git a/.context/research/df-extensions.md b/.context/research/df-extensions.md
index 7470b59..05a3191 100644
--- a/.context/research/df-extensions.md
+++ b/.context/research/df-extensions.md
@@ -73,7 +73,7 @@ TableProviders live) absorbs most of the churn. **Reinforces the
 
 ## ParadeDB — Postgres + DataFusion for search
 
-**Maps to:** §5.10 custom index types, §5.6 mixing engines.
+**Maps to:** §5.4 custom index types (corrected from §5.10 on re-read of MR-737; §5.10 is rank fusion), §5.6 mixing engines.
 
 ParadeDB embeds DataFusion inside Postgres as a query executor for
 analytics workloads. They have a custom `pg_search` extension that
diff --git a/.context/research/duckdb.md b/.context/research/duckdb.md
index 5c947b3..fff712f 100644
--- a/.context/research/duckdb.md
+++ b/.context/research/duckdb.md
@@ -3,8 +3,17 @@
 **Repo:** [`github.com/duckdb/duckdb`](https://github.com/duckdb/duckdb)
 (C++, ~25k★).
 **MR-925 §3.5 mapping:** §5.2 (factorization alternatives — DuckDB
-takes a different approach), §5.6 (vectorized scan), §5.10 (custom
-extensions / index types).
+takes a different approach), §5.6 (vectorized scan), §5.4 (custom
+extensions / index types — corrected on re-read of MR-737; §5.10 is
+rank fusion, *not* index types).
+
+**§-numbering note:** the original MR-925 cross-references in this file said
+§5.10 (custom index types) and §5.11 (multi-statement transactions). On a
+full re-read of MR-737, §5.10 is "First-class scores and rank fusion" and
+§5.11 is "Substrate choice — DataFusion vs. custom executor". The correct
+mapping is §5.4 (custom index plugin model) and §5.12 (mutation IR /
+multi-statement transactions). Section headings below are updated to use the
+corrected numbering.
 **Date:** 2026-05-12.
 
 ---
@@ -51,7 +60,7 @@ For graph workloads with high expansion (per-row neighbor counts >
 workloads vs. a fully-vectorized non-factorized engine. Validates the
 design choice for MR-737.
 
-### Custom extensions / index types — maps to MR-737 §5.10
+### Custom extensions / index types — maps to MR-737 §5.4
 
 DuckDB has a **rich extension API**:
 - `duckdb_extension.h` — public C ABI for loading shared libraries.
@@ -64,13 +73,13 @@ core crate." DuckDB extensions can register completely new index types
 without forking the core. Compare to Lance (§1.2): Lance has the
 necessary trait surface but the registry is `pub(crate)`.
 
-**Decision impact on MR-737 §5.10:** DuckDB demonstrates that the
+**Decision impact on MR-737 §5.4:** DuckDB demonstrates that the
 plugin pattern is workable and prevalent. Our §1.2 blocker (Lance's
 `pub(crate)` registry) is a fixable one — it's not an architectural
 question, just an API-visibility question. Worth a DuckDB-cited
 upstream PR to Lance.
 
-### Multi-statement transactions — maps to MR-737 §5.11
+### Multi-statement transactions — maps to MR-737 §5.12
 
 DuckDB uses MVCC with **per-transaction undo logs** for in-memory
 state. Persistent state is flushed at commit time.
@@ -118,7 +127,7 @@ sink builds and source emits. This is the same shape DataFusion's
 - **§5.6 (scan model):** Pull-based vectorized scan is the right
   shape. DataFusion already gives us this; no new patterns from
   DuckDB.
-- **§5.10 (custom index types):** DuckDB is the proof-of-concept for
+- **§5.4 (custom index plugin model):** DuckDB is the proof-of-concept for
   fully-extensible index plugins. Cite when raising the Lance
   upstream `pub(crate)` issue.
 
diff --git a/.context/research/lancedb.md b/.context/research/lancedb.md
index 2dc86ec..08ba440 100644
--- a/.context/research/lancedb.md
+++ b/.context/research/lancedb.md
@@ -2,8 +2,9 @@
 
 **Repo:** [`github.com/lancedb/lancedb`](https://github.com/lancedb/lancedb)
 (OSS LanceDB, Rust + Python).
-**MR-925 §3.2 mapping:** §5.6 (TableProvider integration), §5.10
-(custom index types), §5.12 (mutation-as-IR), §5.1 (vector search as
+**MR-925 §3.2 mapping:** §5.6 (TableProvider integration), §5.4 (Lance index
+plugin model — *not* §5.10, which is rank fusion; corrected on re-read of MR-737),
+§5.12 (mutation-as-IR), §5.1 + §5.10 (vector search as
 operator vs UDF).
 **Date:** 2026-05-12.
 
@@ -36,7 +37,10 @@ DataFusion's `TableProvider` trait. Key files (Rust side):
   for predicates Lance can fully handle and `::Inexact`/`::Unsupported`
   for the rest. This is the exact pattern §5.6 prescribes.
 
-### Vector search as a query operator — maps to MR-737 §5.1, §5.10
+### Vector search as a query operator — maps to MR-737 §5.1 + §5.10
+
+(§5.1 = unified IR with `VectorSearch` as `IROp`; §5.10 = first-class scores +
+rank fusion; VectorSearch emits `_score`/`_rank` per §5.10.)
 
 LanceDB's `Table::query(vector)` builds a `VectorQuery` and lowers it
 to a Lance scan with a `nearest` filter. **At the DataFusion level**,
@@ -119,7 +123,7 @@ the rest must be designed fresh.
 - **§5.6 (TableProvider integration):** LanceDB's filter-pushdown
   conversion is directly reusable. Capability advertisement beyond
   filters (partitioning, SIP) is OmniGraph-original.
-- **§5.10 (custom index types):** LanceDB integrates with Lance's
+- **§5.4 (Lance index plugin model):** LanceDB integrates with Lance's
   built-in scalar/vector indices but does **not** add custom index
   types from outside the lance crate. Per Experiment 1.2, the plugin
   registry is `pub(crate)`. LanceDB is not a reference for solving
diff --git a/.context/research/trino.md b/.context/research/trino.md
index 51e7080..5d34f97 100644
--- a/.context/research/trino.md
+++ b/.context/research/trino.md
@@ -2,9 +2,20 @@
 
 **Repo:** [`github.com/trinodb/trino`](https://github.com/trinodb/trino)
 (Java, ~10k★).
-**MR-925 §3.6 mapping:** §5.7 (cost-based optimizer at scale), §5.11
-(distributed transaction patterns — informative, NOT prescriptive),
-§5.10 (connector SPI).
+**MR-925 §3.6 mapping:** §5.7 (cost-based optimizer at scale), §5.12
+(distributed transaction patterns — informative, NOT prescriptive; per-table
+branches are §5.12, not §5.11 which is "DataFusion vs custom executor"),
+§5.4 + §5.6 (connector SPI as the plugin/capability surface — corrected on
+re-read of MR-737; §5.10 is rank fusion, *not* connector SPI).
+
+**§-numbering note:** original MR-925 cross-references in this file used
+§5.10 (connector SPI) and §5.11 (per-table txn branches / distributed
+transactions). On a full re-read of MR-737, §5.10 is "First-class scores
+and rank fusion" and §5.11 is "Substrate choice — DataFusion vs. custom
+executor (A)". The correct mapping for the plugin/capability surface is
+§5.4 (Lance index plugin model) + §5.6 (capability-bearing storage trait);
+the correct mapping for per-table txn branches is §5.12 (mutation IR with
+per-table Lance native branches). Section headings below are updated.
 **Date:** 2026-05-12.
 
 ---
@@ -37,7 +48,7 @@ greedy** for larger queries.
   most useful histogram (informs neighbor-expansion cardinality).
   **Add to §5.7 spec.**
 
-### Connector SPI — maps to MR-737 §5.10
+### Connector SPI — maps to MR-737 §5.4 + §5.6
 
 Trino's "Connector SPI" is a Java interface for adding new data
 sources (Postgres, MySQL, Hive, Iceberg, …). The interface is **fully
@@ -55,7 +66,7 @@ public** — connectors are JARs that get registered at server startup.
   `TableProvider` (less stable, breaks ~once per release). Trino is
   the reference for "what a properly-versioned plugin API looks like."
 
-### Distributed transactions — maps to MR-737 §5.11 (informative)
+### Distributed transactions — maps to MR-737 §5.12 + §10 Open Q9 (informative)
 
 Trino is a **stateless query coordinator**: it does NOT manage
 transactions itself. Each connector's underlying system manages its
@@ -64,8 +75,8 @@ identifier that lets multiple queries see the same snapshot of each
 connected system.
 
 **Why this matters:** Trino's approach is the opposite of MR-737
-§5.11 (per-table txn branches with cross-table coordination). Trino
-punts on the cross-system coordination problem; MR-737 doesn't.
+§5.12 (per-table Lance native txn branches with cross-table coordination).
+Trino punts on the cross-system coordination problem; MR-737 doesn't.
 
 **Re-usable pattern:** None directly. Trino is the **anti-reference**
 — what NOT to do if you want cross-table ACID. MR-737's design is
@@ -123,16 +134,17 @@ encoding; we'd be ahead of Trino on this dimension.
   OmniGraph is currently in-process (Phase 0–2). The distributed
   patterns become relevant only at Phase 3+.
 - **The stateless-coordinator approach to transactions.** MR-737
-  §5.11 explicitly wants stateful per-table txn branches.
+  §5.12 explicitly wants stateful per-table txn branches.
 
 ## Decision impact on MR-737
 
 - **§5.7 (cost model):** Add memory cost as third component;
   out-degree histograms as the most-useful additional statistic.
-- **§5.10 (connector SPI / plugin):** Trino is the gold standard for
-  a stable, versioned plugin API. Cite when designing OmniGraph's
-  custom-index-type registration spec.
-- **§5.11 (transactions):** Trino is anti-reference; MR-737 is
+- **§5.4 + §5.6 (connector SPI / plugin / capability surface):** Trino is
+  the gold standard for a stable, versioned plugin API. Cite when designing
+  OmniGraph's custom-index-type registration spec (§5.4) and the storage
+  trait capability surface (§5.6).
+- **§5.12 (transactions):** Trino is anti-reference; MR-737 is
   strictly more ambitious. Document this.
 - **§5.6 (dynamic filters):** Validates the §1.5 design; we'd be
   ahead of Trino with roaring-encoded dynamic filters.