diff --git a/llp/0032-github-llm-graph-bridge.decision.md b/llp/0032-github-llm-graph-bridge.decision.md
index 5fdb572b..a39fbe01 100644
--- a/llp/0032-github-llm-graph-bridge.decision.md
+++ b/llp/0032-github-llm-graph-bridge.decision.md
@@ -6,6 +6,10 @@
 **Author:** Phil / Claude
 **Date:** 2026-06-18
 **Related:** LLP 0012, LLP 0023, LLP 0028, LLP 0030
+**Extended-by:** [LLP 0261](./0261-absent-column-pad-value-per-path.decision.md)
+(#capture's "padding absent physical columns to null" holds only on the
+column-stream path; the row path pads with a cell reading `undefined` post
+LLP 0241, and whether the split collapses is an open question)
 
 > The context graph is a shared substrate for many sources (LLP 0023). When two
 > sources describe the **same real-world entity**, they should land on **one
diff --git a/llp/0261-absent-column-pad-value-per-path.decision.md b/llp/0261-absent-column-pad-value-per-path.decision.md
new file mode 100644
index 00000000..f38a3c0a
--- /dev/null
+++ b/llp/0261-absent-column-pad-value-per-path.decision.md
@@ -0,0 +1,106 @@
+# LLP 0261: The gateway's absent-column pad value is per read path, not "null"
+
+**Type:** Decision
+**Status:** Accepted
+**Systems:** Query, Sources
+**Author:** Phil / Claude
+**Date:** 2026-08-18
+**Related:** LLP 0015, LLP 0055, LLP 0241
+**Extends:** LLP 0032 (#capture: the parenthetical "padding absent physical
+columns to null" holds only on the column-stream path; the row path pads with
+a cell that reads `undefined`)
+
+> Corrects one sentence of [LLP 0032 §Capture](./0032-github-llm-graph-bridge.decision.md#capture)
+> without changing what it decided. The guarantee 0032 needed, a
+> declared-but-physically-absent column stays addressable and never throws,
+> holds on both read paths. The **value** such a read yields is not one thing:
+> `null` on the column stream, `undefined` on the row path. This document
+> records what each path does and the evidence; it deliberately does **not**
+> settle whether the two values should collapse into one.
+
+## Context
+
+<a id="compression"></a>
+
+### What 0032's sentence compresses
+
+LLP 0032 §Capture says the gateway data source "exposes its **declared**
+schema columns (padding absent physical columns to null)". The load-bearing
+half is correct and unchanged: `withSchemaColumns` in
+`hypaware-core/plugins-workspace/ai-gateway/src/dataset.js` advertises the
+dataset's declared schema over partitions that predate part of it, so a
+contract or query naming a new column does not throw `ColumnNotFoundError`
+over a pre-v7 partition. That is all 0032's design (additive nullable columns,
+no partition-label bump, no cache wipe) actually depends on.
+
+The parenthetical "to null" is the compression. It states one pad value where
+the tree has two, one per read path. It is the same compression LLP 0015's
+"Multi-partition union" section once carried ("reads as null, never throws",
+corrected in #731 / PR #740) and that PR #821 uncompressed again after LLP
+0241; after those corrections, this sentence in 0032 was the last remaining
+prose asserting the two paths read the same value (issue #823 item 1).
+
+## What each path actually does
+
+<a id="column-stream"></a>
+
+### Column-stream path: `null`, by normalization
+
+`withSchemaColumns` forwards the `scanColumn` hook so single-column
+aggregates stay on the engine's streaming fast path. A partition that
+physically lacks the requested column surfaces its values as `undefined`
+holes in the chunk, and the wrapper rewrites them to `null` before yielding
+(`wrapped.scanColumn` in
+`hypaware-core/plugins-workspace/ai-gateway/src/dataset.js`, the
+`chunk[i] === undefined` rewrite). Pinned by
+`test/core/ai-gateway-dataset.test.js` ("streams scanColumn with nulls for a
+physically absent column"), which asserts `strictEqual(v, null)` per value:
+"absent column streams null, not undefined". Here 0032's "pads to null" is
+literally true.
+
+<a id="row-path"></a>
+
+### Row path: `undefined`, by padded cell
+
+Post [LLP 0241 §alignment](./0241-scan-rows-carry-advertised-columns.decision.md#alignment),
+a scan's rows carry the column list the scan advertised, and a column the
+partition does not physically hold gets a **padded cell**: `absentCell` in
+`src/core/query/union-source.js` resolves to `undefined`, and the row's
+`resolved` map is left alone, so the padded column is absent from it. A
+projection of that column answers `undefined` (dropped by `JSON.stringify`),
+not `null`, and `IS NULL` filtering works because the engine evaluates SQL
+null semantics over the `undefined` it reads. LLP 0241 chose `undefined`
+deliberately, as the same value the engine already read for a
+declared-but-absent column on the row path pre-alignment (measured in issue
+#778), so that no query that already returned a value changed its answer.
+Pinned by `test/core/star-expansion-drifted-union.test.js` and
+`test/core/union-source.test.js`. Here 0032's "pads to null" is false.
+
+The `withSchemaColumns` doc comment already states the honest form:
+"the exact value a read of it yields depends on the read path (LLP
+0015#multi-partition-union)", as does LLP 0015 itself after PR #740's
+correction. This document brings 0032's prose in line with both.
+
+## Decision
+
+Read LLP 0032 §Capture's parenthetical as: *padding absent physical columns
+so they stay addressable and never throw; the value a read yields depends on
+the read path*, per this document. Nothing 0032 decided changes: the columns
+are nullable, additive, and queryable over old partitions either way, and no
+runtime behavior moves. 0032 is `Active`, so the correction lands as this
+extending document plus a forward-ref on 0032's header, not an in-place
+rewrite.
+
+<a id="split-not-settled"></a>
+
+## Not settled here: the `null`/`undefined` split
+
+Whether the column-stream path (`null` after normalization) and the row path
+(`undefined` padded cell) should collapse to one value remains an **open
+design question**. [LLP 0241](./0241-scan-rows-carry-advertised-columns.decision.md#the-padded-cell-reads-undefined-and-this-decides-nothing-new)
+explicitly declined to settle it ("Whether that split should be collapsed
+remains an open design question, not settled here"), and issue #823 item 5
+records it as open. Consumers must not branch on which of the two values they
+read (the `scanColumn` comment in the ai-gateway dataset forbids exactly
+that). Settling the split needs its own deliberate LLP; this document only
+stops the last remaining prose from asserting the split does not exist.