feat!: mandatory CAS blob storage with streaming I/O (OG-014)

.gitignore

@@ -5,4 +5,5 @@ node_modules/
 coverage/
 CLAUDE.md
 TASKS-DO-NOT-CHECK-IN.md
+EDITORS-REPORT.md
 git-stunts-git-warp-*.tgz

BACKLOG/OG-010-public-api-design-thinking.md

@@ -1,6 +1,6 @@
 # OG-010 — IBM Design Thinking Pass Over Public APIs And README
 
-Status: ACTIVE
+Status: DONE
 
 ## Problem
 

BACKLOG/OG-014-streaming-content-attachments.md

@@ -1,102 +1,114 @@
-# OG-014 — Stream content attachments through git-cas
+# OG-014 — Mandatory CAS blob storage with streaming I/O
 
-Status: QUEUED
+Status: DONE
 
 Legend: Observer Geometry
 
-## Problem
-
-`getContent()` and `getEdgeContent()` currently return full `Uint8Array`
-buffers. That means attachment reads materialize the entire payload in memory
-before user code can process it.
-
-This is fine for small text blobs, but it is the wrong default shape for large
-attachments:
+Design doc: `docs/design/streaming-cas-blob-storage.md`
 
-- the attachment may not fit comfortably in memory
-- the caller cannot decide between buffered read and stream processing
-- builder-facing docs risk teaching attachment reads as eager byte loads
-- the current blob-storage abstraction still forces `retrieve()` to return a
-  full buffer rather than a stream-capable interface
+## Problem
 
-`git-warp` already contains a `CasBlobAdapter` that stores attachments in
-`git-cas` with CDC chunking, but the public attachment path still terminates in
-buffered reads. That leaves the most scalable backend present but not fully
-expressed through the public API.
+Content blob attachments in `git-warp` have two structural problems:
 
-## Why this matters
+### 1. CAS blob storage is opt-in
 
-WARP graphs can legitimately carry attached documents, artifacts, and other
-payloads that are larger than normal graph properties.
+`attachContent()` and `attachEdgeContent()` accept an optional `blobStorage`
+injection. When callers do not provide it, blobs fall through to raw
+`persistence.writeBlob()` — a single unchunked Git object with no CDC
+deduplication, no encryption support, and no streaming restore path.
 
-The API should make the memory tradeoff explicit:
+This means the substrate's chunking, deduplication, and encryption capabilities
+are present but silently bypassed by default. There is no good reason for a
+content blob to skip CAS. Every blob should be chunked.
 
-- buffered reads when you actually want all bytes in memory
-- streaming reads when you want to process incrementally
+### 2. Neither write nor read paths support streaming
 
-That decision should belong to the caller, not be forced by the default
-attachment API shape.
+**Write path**: `attachContent(nodeId, content)` accepts `Uint8Array | string`.
+The caller must buffer the entire payload in memory before handing it to the
+patch builder. `CasBlobAdapter.store()` then wraps that buffer in
+`Readable.from([buf])` — a synthetic stream from an already-buffered payload.
 
-## Current state
+**Read path**: `getContent(nodeId)` returns `Promise<Uint8Array | null>`. The
+full blob is materialized into memory before the caller can process it.
+`CasBlobAdapter.retrieve()` calls `cas.restore()` which buffers internally.
 
-Today the attachment read path is eager:
+`git-cas` already supports streaming on both sides:
+- `cas.store({ source })` accepts any readable/iterable source
+- `cas.restoreStream()` returns `AsyncIterable<Buffer>`
 
-- `getContent()` -> `Promise<Uint8Array|null>`
-- `getEdgeContent()` -> `Promise<Uint8Array|null>`
-- `BlobStoragePort.retrieve()` -> `Promise<Uint8Array>`
-- default Git blob reads go through `readBlob()` and collect the full blob
-- `CasBlobAdapter` can already store attachment content in `git-cas`, but it
-  still restores into one full buffer via `retrieve()`
-- `git-cas` streaming restore is already used in `CasSeekCacheAdapter`, but not
-  yet exposed through attachment reads
+The streaming substrate is there. It is not expressed through the public API.
 
-## Desired outcome
+## Why this matters
 
-Make `git-cas` the first-class streaming attachment path without breaking the
-simple buffered paths.
+WARP graphs can carry attached documents, media, model weights, and other
+payloads that are legitimately large. The API should not force full in-memory
+buffering on either side of the I/O boundary.
 
-Likely shape:
+- Callers writing large content should be able to pipe a stream in
+- Callers reading large content should be able to consume it incrementally
+- Every blob should get CDC chunking and deduplication as a substrate guarantee
+- The decision between buffered and streaming I/O should belong to the caller
 
-- `getContentStream(nodeId)`
-- `getEdgeContentStream(from, to, label)`
-- `BlobStoragePort.retrieveStream(oid)`
-- `CasBlobAdapter.retrieveStream(oid)` backed by `git-cas restoreStream()`
-- a clear default/recommended way to wire `CasBlobAdapter` into `WarpApp.open()`
-  / `WarpCore.open()` for attachment storage
+## Current state
 
-Buffered helpers should remain available for convenience, but they should be
-clearly layered on top of the stream-capable substrate.
+As of `v15.0.1`:
+
+- `BlobStoragePort`: `store(content, options) → Promise<string>`,
+  `retrieve(oid) → Promise<Uint8Array>` — both buffered
+- `CasBlobAdapter`: fully implemented CAS adapter with CDC chunking, optional
+  encryption, backward-compat fallback to raw Git blobs — but only buffered I/O
+- `CasBlobAdapter` is internal (not exported from `index.js`)
+- `PatchBuilderV2.attachContent()`: accepts `Uint8Array | string`, uses
+  `blobStorage.store()` if injected, else raw `persistence.writeBlob()`
+- `getContent()` / `getEdgeContent()`: returns `Promise<Uint8Array | null>`,
+  uses `blobStorage.retrieve()` if injected, else raw `persistence.readBlob()`
+- `WarpApp` and `WarpCore` do not expose content read methods at all
+- `git-cas` streaming (`restoreStream()`) is already used in
+  `CasSeekCacheAdapter` but not in blob reads
+- `InMemoryGraphAdapter` has `writeBlob()`/`readBlob()` for browser/test path
 
-Longer-term, if attachment storage standardizes on `git-cas`, the builder story
-gets cleaner too:
+## Desired outcome
 
-- large attachments become chunked CAS assets
-- reads can stream incrementally
-- dedupe happens below the API surface
-- legacy raw Git blob attachments can remain readable for compatibility
+1. CAS blob storage is mandatory — no fallback to raw `writeBlob()` for content
+2. Write path accepts streaming input and pipes through without buffering
+3. Read path returns a stream the caller can consume incrementally
+4. Buffered convenience methods remain available, layered on top of streams
+5. Browser and in-memory paths still work via a conforming adapter
+6. Legacy raw Git blob attachments remain readable for backward compatibility
 
 ## Acceptance criteria
 
-1. `git-warp` exposes explicit streaming APIs for node and edge attachments.
-2. Callers can choose stream vs buffered read intentionally.
-3. `BlobStoragePort` grows a stream-capable retrieval contract.
-4. `CasBlobAdapter` supports streaming retrieval via `git-cas`.
-5. `git-cas` becomes the recommended path for large attachment storage.
-6. Legacy raw Git blob attachments remain readable for compatibility.
-7. Builder docs explain when to use buffered reads vs streams.
-8. Large attachment reads no longer require full in-memory buffering by
-   default in the stream path.
+1. Every content blob written through `attachContent()` / `attachEdgeContent()`
+   goes through `BlobStoragePort` — no raw `persistence.writeBlob()` fallback.
+2. `attachContent()` / `attachEdgeContent()` accept streaming input
+   (`AsyncIterable<Uint8Array>`, `ReadableStream`, `Uint8Array`, `string`).
+3. New `getContentStream()` / `getEdgeContentStream()` return
+   `AsyncIterable<Uint8Array>` for incremental consumption.
+4. Existing `getContent()` / `getEdgeContent()` remain as buffered convenience,
+   implemented on top of the stream primitive.
+5. `BlobStoragePort` grows `storeStream()` and `retrieveStream()` methods.
+6. `CasBlobAdapter` implements streaming via `git-cas` natively.
+7. An `InMemoryBlobStorageAdapter` implements the port contract for browser and
+   test paths.
+8. Legacy raw Git blob attachments remain readable through backward-compat
+   fallback in `CasBlobAdapter.retrieveStream()`.
+9. Content stream methods are exposed on `WarpApp` and `WarpCore`.
 
 ## Non-goals
 
-- no automatic conversion of all existing attachment reads to streams
-- no silent breaking change to `getContent()` / `getEdgeContent()`
-- no attempt to solve whole-state out-of-core replay here
+- No automatic migration of existing raw Git blobs to CAS format
+- No silent breaking change to existing `getContent()` / `getEdgeContent()`
+  return types
+- No attempt to solve whole-state out-of-core replay (that is OG-013)
+- No encryption-by-default (encryption remains an opt-in CAS capability)
 
 ## Notes
 
-This item is related to, but narrower than,
-`OG-013-out-of-core-materialization-and-streaming-reads.md`.
-`OG-013` is about whole-state and replay architecture.
-This item is specifically about attachment payload I/O and making
-`git-cas` the streaming/chunked attachment path.
+This item supersedes the original OG-014 scope, which covered only streaming
+reads. The expanded scope now includes mandatory CAS and streaming writes.
+
+Related items:
+- `OG-013`: out-of-core materialization and streaming reads (broader, separate)
+- `B160`: blob attachments via CAS (done, but opt-in — this item makes it
+  mandatory)
+- `B163`: streaming restore for seek cache (done, pattern to follow for blobs)

BACKLOG/OG-015-jsr-documentation-quality.md

@@ -1,6 +1,6 @@
 # OG-015 — Raise JSR documentation quality score
 
-Status: QUEUED
+Status: DONE
 
 Legend: Observer Geometry
 

BACKLOG/README.md

@@ -1,6 +1,6 @@
 # BACKLOG — Observer Geometry
 
-Last updated: 2026-03-28
+Last updated: 2026-03-29
 
 This directory holds promotable pre-design items for the current Observer
 Geometry tranche.
@@ -26,9 +26,9 @@ Workflow:
 | DONE | OG-007 | Expand hash-stability coverage across snapshot flavors | [OG-007-hash-stability-coverage.md](OG-007-hash-stability-coverage.md) |
 | DONE | OG-008 | Make retargeting compatibility a hard major-version cut | [OG-008-retargeting-compatibility.md](OG-008-retargeting-compatibility.md) |
 | QUEUED | OG-009 | Align playback-head and TTD consumers after read nouns stabilize | [OG-009-playback-head-alignment.md](OG-009-playback-head-alignment.md) |
-| ACTIVE | OG-010 | IBM Design Thinking pass over public APIs and README | [OG-010-public-api-design-thinking.md](OG-010-public-api-design-thinking.md) |
+| DONE | OG-010 | IBM Design Thinking pass over public APIs and README | [OG-010-public-api-design-thinking.md](OG-010-public-api-design-thinking.md) |
 | QUEUED | OG-011 | Publish a public API catalog and browser documentation playground | [OG-011-public-api-catalog-and-playground.md](OG-011-public-api-catalog-and-playground.md) |
 | DONE | OG-012 | Audit and reconcile the documentation corpus before v15 | [OG-012-documentation-corpus-audit.md](OG-012-documentation-corpus-audit.md) |
 | QUEUED | OG-013 | Design out-of-core materialization and streaming reads | [OG-013-out-of-core-materialization-and-streaming-reads.md](OG-013-out-of-core-materialization-and-streaming-reads.md) |
-| QUEUED | OG-014 | Stream content attachments through `git-cas` | [OG-014-streaming-content-attachments.md](OG-014-streaming-content-attachments.md) |
-| QUEUED | OG-015 | Raise JSR documentation quality score | [OG-015-jsr-documentation-quality.md](OG-015-jsr-documentation-quality.md) |
+| DONE | OG-014 | Mandatory CAS blob storage with streaming I/O | [OG-014-streaming-content-attachments.md](OG-014-streaming-content-attachments.md) |
+| DONE | OG-015 | Raise JSR documentation quality score | [OG-015-jsr-documentation-quality.md](OG-015-jsr-documentation-quality.md) |

CHANGELOG.md

@@ -5,6 +5,25 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [16.0.0] — 2026-03-29
+
+### Added
+
+- **Streaming content attachment I/O (OG-014)** — `getContentStream()` and `getEdgeContentStream()` return `AsyncIterable<Uint8Array>` for incremental consumption of large content blobs. `attachContent()` and `attachEdgeContent()` now accept `AsyncIterable<Uint8Array>`, `ReadableStream<Uint8Array>`, `Uint8Array`, or `string` — streaming inputs are piped directly to blob storage without intermediate buffering.
+- **`InMemoryBlobStorageAdapter`** — new domain-local adapter implementing `BlobStoragePort` with content-addressed `Map`-based storage for browser and test paths. Exported from the package surface.
+- **`BlobStoragePort.storeStream()` / `retrieveStream()`** — streaming variants of the blob storage port contract. `storeStream()` accepts `AsyncIterable<Uint8Array>`, `retrieveStream()` returns `AsyncIterable<Uint8Array>`.
+- **Content methods on `WarpApp` and `WarpCore`** — `getContent()`, `getContentStream()`, `getContentOid()`, `getContentMeta()` and their edge equivalents are now exposed on both public API surfaces (previously only on `WarpRuntime`).
+
+### Changed
+
+- **CAS blob storage is now mandatory for content attachments** — `attachContent()` and `attachEdgeContent()` always route through `BlobStoragePort`. The raw `persistence.writeBlob()` fallback has been removed. `WarpRuntime.open()` auto-constructs `CasBlobAdapter` for Git-backed persistence (when `plumbing` is available) or `InMemoryBlobStorageAdapter` otherwise.
+- **Content blob tree entries use tree mode** — patch commit trees and checkpoint trees now reference content blobs as `040000 tree` entries (matching CAS tree OIDs) instead of `100644 blob` entries.
+
+### Removed
+
+- **`TraversalService`** — deprecated alias removed. Use `CommitDagTraversalService` directly.
+- **`createWriter()`** — deprecated method removed from `WarpApp` and `WarpCore`. Use `writer()` or `writer(id)` instead.
+
 ## [15.0.0] — 2026-03-28
 
 ## [15.0.1] — 2026-03-28

MIGRATING.md

@@ -0,0 +1,108 @@
+# Migrating to v16
+
+This guide covers the breaking changes in v16.0.0 and how to update your code.
+
+## Content attachments now require blob storage (OG-014)
+
+**What changed:** `attachContent()` and `attachEdgeContent()` no longer fall
+back to raw `persistence.writeBlob()`. They always route through
+`BlobStoragePort`. Without blob storage, they throw `NO_BLOB_STORAGE`.
+
+**Who is affected:** Only consumers who construct `PatchBuilderV2` directly
+(bypassing `WarpRuntime.open()`) without passing `blobStorage`. If you use
+`WarpApp.open()`, `WarpCore.open()`, or `WarpRuntime.open()`, blob storage
+is auto-constructed — no code changes needed.
+
+**How to migrate:**
+
+```javascript
+import InMemoryBlobStorageAdapter from '@git-stunts/git-warp/defaultBlobStorage';
+
+// If you construct PatchBuilderV2 directly, add blobStorage:
+const builder = new PatchBuilderV2({
+  persistence,
+  writerId: 'alice',
+  blobStorage: new InMemoryBlobStorageAdapter(), // or CasBlobAdapter for Git
+  // ...other options
+});
+```
+
+## Streaming content I/O
+
+**What changed:** `attachContent()` and `attachEdgeContent()` now accept
+streaming input (`AsyncIterable<Uint8Array>`, `ReadableStream<Uint8Array>`)
+in addition to `Uint8Array` and `string`. New `getContentStream()` and
+`getEdgeContentStream()` methods return `AsyncIterable<Uint8Array>`.
+
+**Who is affected:** No one — this is additive. Existing code continues to
+work. New streaming APIs are opt-in.
+
+**How to use:**
+
+```javascript
+// Streaming write — pipe a file directly
+import { createReadStream } from 'node:fs';
+const patch = await app.createPatch();
+patch.addNode('doc:1');
+await patch.attachContent('doc:1', createReadStream('large-file.bin'), {
+  size: fileStat.size,
+  mime: 'application/octet-stream',
+});
+await patch.commit();
+
+// Streaming read — consume incrementally
+const stream = await app.getContentStream('doc:1');
+if (stream) {
+  for await (const chunk of stream) {
+    process.stdout.write(chunk);
+  }
+}
+
+// Buffered read — unchanged
+const buf = await app.getContent('doc:1');
+```
+
+## Content blob tree entries use tree mode
+
+**What changed:** Patch commit trees and checkpoint trees now reference
+content blobs as `040000 tree` entries (CAS tree OIDs) instead of
+`100644 blob` entries.
+
+**Who is affected:** Consumers who parse raw Git commit trees and expect
+content anchor entries to use blob mode. This does not affect any public
+API — it is an internal storage format change.
+
+**How to migrate:** If you parse `_content_<oid>` entries from commit trees,
+update your parser to accept `040000 tree` mode.
+
+## `TraversalService` removed
+
+**What changed:** The `TraversalService` export was a deprecated alias for
+`CommitDagTraversalService`. It has been removed.
+
+**How to migrate:**
+
+```javascript
+// Before
+import { TraversalService } from '@git-stunts/git-warp';
+
+// After
+import { CommitDagTraversalService } from '@git-stunts/git-warp';
+```
+
+## `createWriter()` removed
+
+**What changed:** The `createWriter()` method on `WarpApp` was deprecated in
+v15 and has been removed. Use `writer()` instead.
+
+**How to migrate:**
+
+```javascript
+// Before
+const w = await app.createWriter();
+const w2 = await app.createWriter({ persist: 'config', alias: 'secondary' });
+
+// After
+const w = await app.writer();          // resolves from git config or generates
+const w2 = await app.writer('secondary'); // explicit ID
+```