diff --git a/docs/redesign/m2-atomic-reindex-migration-note.md b/docs/redesign/m2-atomic-reindex-migration-note.md
new file mode 100644
index 0000000000..0e5660a7ef
--- /dev/null
+++ b/docs/redesign/m2-atomic-reindex-migration-note.md
@@ -0,0 +1,55 @@
+# M2 Atomic Reindex Migration Note
+
+## Goal
+
+Make reindex completion atomic so search, evaluation, telemetry, and document views never observe a partially replaced index generation.
+
+This branch is intentionally docs-only before implementation. Code changes should start only after this note is reviewed.
+
+## Current Risk
+
+The existing reindex path can delete or replace index artifacts before the replacement generation is fully available. A failure between removal and completion can leave a document with missing chunks, stale retrieval metadata, or inconsistent `index_generation_id` references.
+
+## API And Config Touchpoints
+
+- `src/app/api/documents/[id]/reindex/route.ts`: single-document reindex entrypoint and response contract.
+- `src/app/api/documents/bulk/reindex/route.ts`: bulk reindex request shape and partial-failure reporting.
+- `src/lib/reindex-pipeline.ts`: orchestration boundary for extraction, chunking, embeddings, artifact writes, and final status.
+- `src/lib/ingestion.ts` and `worker/main.ts`: shared indexing behavior that must preserve compatibility with upload and background ingestion.
+- `supabase/functions/indexing-v3-agent/index.ts` and `supabase/functions/indexing-v3-agent/behavior.ts`: remote worker behavior and retry/claim semantics.
+- Supabase tables: `documents`, `document_chunks`, `document_index_units`, `document_memory_cards`, `document_table_facts`, `document_embedding_fields`, and related visual/image tables.
+- Supabase RPCs/functions: `match_document_chunks`, `match_document_chunks_hybrid`, `match_document_chunks_text`, `match_documents_for_query`, `claim_indexing_v3_agent_jobs`, and indexing completion/failure RPCs.
+- Config/env: Supabase project ref `sjrfecxgysukkwxsowpy`, storage buckets, `OPENAI_API_KEY`, worker retry settings, and local no-auth owner defaults used by eval scripts.
+
+## Compatibility Assumptions
+
+- Existing public API payloads remain backwards compatible: reindex routes may add fields, but must not remove current fields or change success/failure status semantics.
+- Existing search and eval flows continue to accept documents without a new generation until the atomic cutover completes.
+- Retrieval RPCs must ignore staged, incomplete generations by default.
+- Existing telemetry keys remain valid. New generation metadata can be additive.
+- Rollback must preserve the last completed generation until the new generation is fully committed.
+
+## Proposed Migration Shape
+
+1. Add a durable generation record or equivalent metadata state for `pending`, `committing`, `completed`, and `failed`.
+2. Write all replacement artifacts with a new generation id while leaving the previous completed generation readable.
+3. Switch the document's active generation in one database transaction or RPC after all required artifacts validate.
+4. Delete or mark stale artifacts only after the active-generation switch succeeds.
+5. Update retrieval RPC filters so default reads use only the active completed generation.
+6. Add repair/recovery logic for stuck `pending` or `committing` generations.
+
+## Required Tests
+
+- Unit tests for generation-state transitions and rollback behavior.
+- Route tests for single and bulk reindex response compatibility.
+- Worker tests for retry after extraction, embedding, and DB-write failures.
+- Retrieval tests proving staged generations are invisible until commit.
+- Eval/search smoke tests proving existing telemetry and source links remain compatible.
+
+## Acceptance Criteria
+
+- A failed reindex leaves the previous searchable generation intact.
+- A successful reindex switches all retrieval surfaces to one completed generation.
+- No search result mixes chunks from old and new generations for the same document.
+- Recovery tooling can identify and clean abandoned staged generations.
+- `npm run verify:cheap` passes, plus the smallest relevant retrieval/indexing checks.