diff --git a/docs/branch-review-records/3d45d96c0c8e97d835c1a3f6c12b320255c0164dc545dce172e67703b2be6fda.record.md b/docs/branch-review-records/3d45d96c0c8e97d835c1a3f6c12b320255c0164dc545dce172e67703b2be6fda.record.md
new file mode 100644
index 0000000000..9a3057fb96
--- /dev/null
+++ b/docs/branch-review-records/3d45d96c0c8e97d835c1a3f6c12b320255c0164dc545dce172e67703b2be6fda.record.md
@@ -0,0 +1 @@
+| 2026-08-14 | claude/rag-zod-hardening-tranche1 | 1fb98a174e815ce2d7e105d9339d766d3f54b02b | PR #1946 review-and-fix | fixed PR-introduced maintainability blocker; strengthened retrieval provenance and visual shape validation; merged latest main | offline: git diff --check; Prettier changed files; check:maintainability-budgets; branch-review-ledger; ledger-write-discipline; docs:check-links; CI pending |
diff --git a/docs/branch-review-records/bb4cf300c0917336cbfb205dc3631242f4e32aab35d3d3bfe1d0654c41f7bf34.record.md b/docs/branch-review-records/bb4cf300c0917336cbfb205dc3631242f4e32aab35d3d3bfe1d0654c41f7bf34.record.md
new file mode 100644
index 0000000000..c0c42f5a5d
--- /dev/null
+++ b/docs/branch-review-records/bb4cf300c0917336cbfb205dc3631242f4e32aab35d3d3bfe1d0654c41f7bf34.record.md
@@ -0,0 +1 @@
+| 2026-08-14 | claude/rag-zod-hardening-tranche1 | 0fded0878b6f8e35ef50b8c97a63af6e78e1f716 | rag.ts retrieval RPC row Zod shape contract (ledger #212 tranche 1) | Approved — 4 unchecked as SearchResult[] casts on retrieval RPC rows replaced with assertRetrievalRows; no ranking/ordering/scoring logic touched | verify:pr-local (all steps green except pre-existing check:medication-lexicon-report staleness on main); test 602 files/6517 passed; typecheck exit 0; eval:rag:offline 36 golden cases 23 suites, 579 passed |
diff --git a/src/lib/rag/rag-row-contracts.ts b/src/lib/rag/rag-row-contracts.ts
new file mode 100644
index 0000000000..566ce5cf53
--- /dev/null
+++ b/src/lib/rag/rag-row-contracts.ts
@@ -0,0 +1,129 @@
+import { z } from "zod";
+import { logger } from "@/lib/logger";
+import { normalizeOptionalSourceMetadata } from "@/lib/source-metadata";
+import type { SearchResult } from "@/lib/types";
+
+/**
+ * Runtime shape contract for a row returned by a retrieval RPC.
+ *
+ * Retrieval rows are untrusted external data, not a compile-time guarantee. The RPCs are
+ * versioned (`*_v2` with a legacy fallback) and the live database has been observed to
+ * drift from the migrations in this repo — `docs/outstanding-issues.md` `#316` records ten
+ * retrieval RPC bodies diverging, with weekly live-drift red since 2026-07-26. Before this
+ * contract existed, `rag.ts` asserted those rows straight into the ranking pipeline with a
+ * bare `as SearchResult[]`, so a renamed column or a numeric returned as a string did not
+ * fail — it misranked or mis-cited silently, which is the worst failure mode for a clinical
+ * reference surface.
+ *
+ * The schema is deliberately asymmetric:
+ *
+ * - **Strict on the ranking, citation, and evidence fields.** The required chunk identity,
+ *   provenance, and visual fields are `not null` in `supabase/schema.sql`, so requiring them
+ *   cannot reject a row that works today. The four score fields are `.nullish()` — absent or
+ *   null already flows through the downstream `?? 0` handling unchanged — but a *string where
+ *   a number belongs* is rejected, which is precisely the silent-misranking case this exists
+ *   to catch.
+ * - **Loose about everything else.** Column sets genuinely differ between RPC versions
+ *   (`retrieval_synopsis` is absent from the older base hybrid function; `document_labels`
+ *   and `document_summary` only appear on `match_document_chunks_v2`). `z.looseObject`
+ *   preserves unknown keys rather than stripping them, so a harmless schema difference
+ *   never becomes an outage or silent data loss.
+ */
+const retrievalImageSchema = z.looseObject({
+  id: z.string().min(1),
+  page_number: z.number().int().nullable(),
+  storage_path: z.string(),
+  caption: z.string(),
+});
+
+const retrievalRowSchema = z.looseObject({
+  id: z.string().min(1),
+  document_id: z.string().min(1),
+  title: z.string(),
+  file_name: z.string(),
+  page_number: z.number().int().nullable(),
+  chunk_index: z.number().int(),
+  section_heading: z.string().nullable(),
+  content: z.string(),
+  image_ids: z.array(z.string()),
+  source_metadata: z.record(z.string(), z.unknown()).nullable(),
+  images: z.array(retrievalImageSchema),
+  similarity: z.number().nullish(),
+  text_rank: z.number().nullish(),
+  hybrid_score: z.number().nullish(),
+  rrf_score: z.number().nullish(),
+});
+
+const retrievalRowsSchema = z.array(retrievalRowSchema);
+
+/** Cap on reported issues; a wholesale shape change would otherwise report one per row. */
+const MAX_REPORTED_ISSUES = 5;
+
+/**
+ * Thrown when a retrieval RPC returns rows that do not satisfy the ranking contract.
+ *
+ * The message carries only Zod issue paths and codes — never a row value. Retrieval rows
+ * contain clinical document text, so echoing one into a log or an error response would
+ * leak source content past the privacy boundary that `query-privacy.ts` maintains.
+ */
+export class RetrievalRowShapeError extends Error {
+  readonly rpc: string;
+  readonly issues: string[];
+
+  constructor(rpc: string, issues: string[]) {
+    super(`Retrieval RPC "${rpc}" returned rows that do not match the ranking contract: ${issues.join("; ")}`);
+    this.name = "RetrievalRowShapeError";
+    this.rpc = rpc;
+    this.issues = issues;
+  }
+}
+
+function describeIssues(error: z.ZodError): string[] {
+  const described = error.issues
+    .slice(0, MAX_REPORTED_ISSUES)
+    .map((issue) => `${issue.path.join(".") || "<root>"}: ${issue.message}`);
+  const remaining = error.issues.length - described.length;
+  return remaining > 0 ? [...described, `and ${remaining} more`] : described;
+}
+
+/**
+ * Validate retrieval RPC rows before they enter the ranking pipeline.
+ *
+ * This asserts rather than transforms: on success the caller keeps the original array and
+ * row objects, so object identity, key order, and nested `images` / `source_metadata`
+ * references are unchanged from what the RPC returned. That matters because ranking is a
+ * live-validated protected surface (`docs/rag-behaviour/`) — validation must be observable
+ * only when the data is already wrong.
+ *
+ * Logs before throwing so drift stays visible even where a caller degrades: the vector
+ * fallback in `rag.ts` catches retrieval failures and falls back to lexical results when it
+ * has them, which would otherwise swallow the signal entirely.
+ */
+export function assertRetrievalRows(rows: unknown, rpc: string): asserts rows is SearchResult[] {
+  const parsed = retrievalRowsSchema.safeParse(rows);
+  if (parsed.success) return;
+  const issues = describeIssues(parsed.error);
+  logger.error("retrieval_row_shape_mismatch", {
+    rpc,
+    issues,
+    rowCount: Array.isArray(rows) ? rows.length : null,
+  });
+  throw new RetrievalRowShapeError(rpc, issues);
+}
+
+/** Build and validate the locally retrieved rows used as document-summary context. */
+export function buildDocumentSummaryResults(
+  chunks: unknown[],
+  document: { title: string; file_name: string; metadata?: unknown },
+): SearchResult[] {
+  const results = chunks.map((chunk) => ({
+    ...(chunk as object),
+    title: document.title,
+    file_name: document.file_name,
+    source_metadata: normalizeOptionalSourceMetadata(document.metadata),
+    similarity: 1,
+    images: [],
+  }));
+  assertRetrievalRows(results, "document_summary_context");
+  return results;
+}
diff --git a/src/lib/rag/rag.ts b/src/lib/rag/rag.ts
index e09f968c1f..e16eab72e1 100644
--- a/src/lib/rag/rag.ts
+++ b/src/lib/rag/rag.ts
@@ -1,6 +1,7 @@
 import { createAdminClient } from "@/lib/supabase/admin";
 import { loadDocumentSummaryContext } from "@/lib/rag/rag-document-summary-context";
 import { generationFailureDetailToken } from "@/lib/rag/rag-generation-failure-diagnostics";
+import { assertRetrievalRows, buildDocumentSummaryResults } from "@/lib/rag/rag-row-contracts";
 import { answerInstructions } from "@/lib/rag/rag-answer-instructions";
 import { retrievalAccessScopeForArgs, retrievalRpcScopeArgs } from "@/lib/owner-scope";
 import {
@@ -2179,13 +2180,17 @@ export async function searchChunksWithTelemetry(
 
   const { data: hybridData, error: hybridError } = hybridResult;
   if (hybridError) recordHybridRpcError(telemetry, "match_document_chunks_hybrid", hybridError);
-  telemetry.vector_candidate_count = hybridData?.length ?? 0;
-  recordRetrievalLayer(telemetry, "hybrid_vector", hybridData?.length ?? 0, {
+  // On a hybrid RPC error `hybridData` is null, so this validates an empty array and the
+  // existing hybrid-error -> vector-fallback path below is reached unchanged.
+  const hybridRows = hybridData ?? [];
+  assertRetrievalRows(hybridRows, "match_document_chunks_hybrid");
+  telemetry.vector_candidate_count = hybridRows.length;
+  recordRetrievalLayer(telemetry, "hybrid_vector", hybridRows.length, {
     latencyMs: hybridResult.latencyMs,
-    topScore: layerTopScore((hybridData ?? []) as SearchResult[]),
+    topScore: layerTopScore(hybridRows),
   });
   const vectorCandidates = mergeSearchResults(
-    mergeSearchResults((hybridData ?? []) as SearchResult[], embeddingFieldCandidates),
+    mergeSearchResults(hybridRows, embeddingFieldCandidates),
     indexUnitCandidates,
   );
 
@@ -2257,7 +2262,9 @@ export async function searchChunksWithTelemetry(
       );
 
       if (error) throw new Error(error.message);
-      return (data ?? []) as SearchResult[];
+      const rows = data ?? [];
+      assertRetrievalRows(rows, "match_document_chunks");
+      return rows;
     }),
   ).catch((error) => {
     if (!args.forceEmbedding && textFastResults.length > 0) return [] as SearchResult[][];
@@ -4304,15 +4311,7 @@ export async function summarizeDocument(documentId: string, ownerId?: string, op
     } satisfies RagAnswer;
   }
 
-  const documentMetadata = (document as { metadata?: unknown }).metadata;
-  const results = committedChunks.map((chunk) => ({
-    ...chunk,
-    title: document.title,
-    file_name: document.file_name,
-    source_metadata: normalizeOptionalSourceMetadata(documentMetadata),
-    similarity: 1,
-    images: [],
-  })) as SearchResult[];
+  const results = buildDocumentSummaryResults(committedChunks, document);
 
   const summaryInstructions = `Summarize a clinical document for practical psychiatric use in Perth, Australia.
 Use only the excerpts provided. Use a layered response: make the answer field a plain high-yield clinical paragraph,
diff --git a/tests/rag-retrieval-row-contract.test.ts b/tests/rag-retrieval-row-contract.test.ts
new file mode 100644
index 0000000000..4aa447865f
--- /dev/null
+++ b/tests/rag-retrieval-row-contract.test.ts
@@ -0,0 +1,146 @@
+import { describe, expect, it, vi } from "vitest";
+import { RetrievalRowShapeError, assertRetrievalRows, buildDocumentSummaryResults } from "@/lib/rag/rag-row-contracts";
+
+vi.mock("@/lib/logger", () => ({
+  logger: { debug: vi.fn(), info: vi.fn(), warn: vi.fn(), error: vi.fn() },
+}));
+
+// A realistic `match_document_chunks_hybrid_v2` row. Column list mirrors the RPC's
+// `returns table (...)` in supabase/migrations/20260713020000_owner_plus_public_retrieval.sql.
+/** A row with one column dropped, standing in for an RPC whose shape has drifted. */
+function withoutColumn(column: string, overrides: Record<string, unknown> = {}) {
+  const row: Record<string, unknown> = hybridRow(overrides);
+  delete row[column];
+  return row;
+}
+
+function hybridRow(overrides: Record<string, unknown> = {}) {
+  return {
+    id: "3f1a2b6c-1111-4aaa-8bbb-000000000001",
+    document_id: "3f1a2b6c-2222-4aaa-8bbb-000000000002",
+    title: "RANZCP Mood Disorders Guideline",
+    file_name: "ranzcp-mood.pdf",
+    page_number: 14,
+    chunk_index: 7,
+    section_heading: "Lithium monitoring",
+    content: "Check serum lithium 5 days after any dose change.",
+    retrieval_synopsis: null,
+    image_ids: [],
+    source_metadata: { document_status: "current" },
+    similarity: 0.91,
+    text_rank: 0.42,
+    hybrid_score: 0.88,
+    rrf_score: 0.031,
+    images: [],
+    ...overrides,
+  };
+}
+
+describe("retrieval row shape contract", () => {
+  it("accepts a realistic hybrid RPC row without mutating it", () => {
+    const rows: unknown = [hybridRow()];
+    const before = structuredClone(rows);
+    const firstRowReference = (rows as unknown[])[0];
+
+    expect(() => assertRetrievalRows(rows, "match_document_chunks_hybrid")).not.toThrow();
+
+    // Assert, do not transform: ranking is a live-validated protected surface, so a valid
+    // row must reach it byte-identical and by the same reference.
+    expect(rows).toEqual(before);
+    expect((rows as unknown[])[0]).toBe(firstRowReference);
+  });
+
+  it("builds valid document-summary rows outside the retrieval monolith", () => {
+    const chunk = hybridRow({ title: "stale", file_name: "stale.pdf", similarity: 0.2 });
+
+    const rows = buildDocumentSummaryResults([chunk], {
+      title: "Current title",
+      file_name: "current.pdf",
+      metadata: { document_status: "current" },
+    });
+
+    expect(rows[0]).toMatchObject({
+      id: chunk.id,
+      title: "Current title",
+      file_name: "current.pdf",
+      similarity: 1,
+      source_metadata: expect.objectContaining({ document_status: "current" }),
+    });
+  });
+
+  it("preserves unknown columns so an RPC version difference is not data loss", () => {
+    const rows: unknown = [hybridRow({ document_labels: [{ id: "l1" }], a_future_column: 42 })];
+
+    assertRetrievalRows(rows, "match_document_chunks_v2");
+
+    expect(rows[0]).toMatchObject({ document_labels: [{ id: "l1" }], a_future_column: 42 });
+  });
+
+  it("rejects a numeric score returned as a string", () => {
+    const rows: unknown = [hybridRow({ similarity: "0.91" })];
+
+    expect(() => assertRetrievalRows(rows, "match_document_chunks_hybrid")).toThrow(RetrievalRowShapeError);
+  });
+
+  it("rejects a row missing its chunk identity", () => {
+    const rows: unknown = [withoutColumn("id")];
+
+    expect(() => assertRetrievalRows(rows, "match_document_chunks_hybrid")).toThrow(RetrievalRowShapeError);
+    expect(() => assertRetrievalRows([hybridRow({ document_id: "" })], "x")).toThrow(RetrievalRowShapeError);
+  });
+
+  it("rejects malformed provenance and visual fields before they can miscite or crash rendering", () => {
+    expect(() => assertRetrievalRows([withoutColumn("title")], "match_document_chunks_hybrid")).toThrow(
+      RetrievalRowShapeError,
+    );
+    expect(() =>
+      assertRetrievalRows([hybridRow({ image_ids: "not-an-array" })], "match_document_chunks_hybrid"),
+    ).toThrow(RetrievalRowShapeError);
+    expect(() => assertRetrievalRows([hybridRow({ images: [{}] })], "match_document_chunks_hybrid")).toThrow(
+      RetrievalRowShapeError,
+    );
+  });
+
+  it("accepts absent or null scores, which downstream already coalesces to 0", () => {
+    expect(() => assertRetrievalRows([withoutColumn("text_rank")], "match_document_chunks")).not.toThrow();
+    expect(() => assertRetrievalRows([hybridRow({ rrf_score: null })], "match_document_chunks")).not.toThrow();
+    expect(() => assertRetrievalRows([], "match_document_chunks")).not.toThrow();
+  });
+
+  it("names the RPC and leaks no row content in the error", () => {
+    const secret = "Check serum lithium 5 days after any dose change.";
+    let thrown: RetrievalRowShapeError | null = null;
+    try {
+      assertRetrievalRows([hybridRow({ hybrid_score: "0.88" })], "match_document_chunks_hybrid");
+    } catch (error) {
+      thrown = error as RetrievalRowShapeError;
+    }
+
+    expect(thrown).toBeInstanceOf(RetrievalRowShapeError);
+    expect(thrown?.rpc).toBe("match_document_chunks_hybrid");
+    expect(thrown?.message).toContain("match_document_chunks_hybrid");
+    expect(thrown?.message).toContain("hybrid_score");
+    // Retrieval rows carry clinical document text; it must never reach a log or a response.
+    expect(thrown?.message).not.toContain(secret);
+    expect(thrown?.message).not.toContain("ranzcp-mood.pdf");
+  });
+
+  it("caps the reported issues so a wholesale shape change stays readable", () => {
+    const rows = Array.from({ length: 20 }, () => hybridRow({ similarity: "nope" }));
+
+    let thrown: RetrievalRowShapeError | null = null;
+    try {
+      assertRetrievalRows(rows, "match_document_chunks_hybrid");
+    } catch (error) {
+      thrown = error as RetrievalRowShapeError;
+    }
+
+    expect(thrown?.issues).toHaveLength(6);
+    expect(thrown?.issues.at(-1)).toBe("and 15 more");
+  });
+
+  it("rejects a payload that is not an array of rows", () => {
+    expect(() => assertRetrievalRows({ rows: [] }, "match_document_chunks_hybrid")).toThrow(RetrievalRowShapeError);
+    expect(() => assertRetrievalRows(null, "match_document_chunks_hybrid")).toThrow(RetrievalRowShapeError);
+  });
+});