v4.1.3: edge profile through persistence backend + corpus hardening

Squashed v4.1 release train. Closes the v4.0 GA "edge profile is
theatrical" gap (handlers held a 6-row sqlite-vec smoke harness while
list/get/search/create/update/delete still acquired raw asyncpg
connections from the Postgres pool) and ships the v4.1.1-1.3 hardening
on top.

## Persistence-backend dispatch (was v4.1.0 / v4.1.1)

Every handler that previously did `_pool.acquire()` now goes through
`_persistence_backend.transactional()` so the active backend
(PostgresBackend or SqliteBackend) decides how to access storage:

- list / get / search / create / update / delete memories
- bulk_create_memories
- /stats (MemoryStatsRow + CompressionStatsRow via gather_stats)
- consultation completion + audit log + DAG live-merge webhook fan-out
  now go through the transactional outbox (delivery scheduled
  post-commit, not mid-handler)

VisibilityFilter (ROOT_BYPASS / READABLE / OWN_ONLY) replaces the
ad-hoc namespace + permission_mode predicates that handlers used to
inline. Fail-closed: namespace=None for non-root callers raises at
filter construction.

`PostgresBackend.transactional()` uses the canonical `async with
self._pool.acquire() as conn:` pattern (the previous shape awaited
acquire() then async-with on the resolved proxy and broke on
asyncpg's PoolConnectionProxy).

## GRAEAE provider routing (was v4.1.2)

- `_is_reasoning_variant` helper in domain/graeae/engine.py prefers
  -non-reasoning over -reasoning siblings when ranking model
  candidates (xai grok-4.20-0309-non-reasoning was being shadowed by
  the -reasoning variant on len-tiebreak; surfaced as `\confidence{N}`
  trailers on consensus output).
- Reasoning-variant rows in db/migrations_*model_registry.sql marked
  deprecated; non-reasoning kept canonical.
- Container env standard codified: `MNEMOS_KEYS_PATH` env +
  `~/.api_keys_master.json:/etc/mnemos/api_keys.json:ro` mount +
  `RATE_LIMIT_STORAGE=memory://` are mandatory in every `mnemos serve`
  container or GRAEAE silently 401s every consultation.

## Corpus-review hardening (was v4.1.3)

Fixes the 10 issues from the v4.1.1 full-corpus Codex review
(docs/CORPUS-REVIEW-2026-04-29.md).

Security:
- Pin webhook delivery DNS resolution from validation through HTTP
  connect via `PinnedDNSAsyncHTTPTransport` (closes DNS-rebinding
  SSRF).
- Apply the same SSRF URL validation to federation peers. Private
  peer URLs require `FEDERATION_ALLOW_PRIVATE=true`.

Reliability:
- Release GRAEAE provider concurrency slots in finally during
  cancelled fan-out (no slot leak on client disconnect).

Profile correctness:
- Sessions, entities, state, MORPHEUS routes now return explicit 503
  on edge (SQLite) profile via `_require_postgres_backend`.
- MORPHEUS run telemetry restricted to root/operator callers.
- DAG read preflight aligned with memory read visibility; branch +
  merge writes stay strict-owner-scoped.
- Typed `AuthSettings` + server-profile fail-closed auth defaults via
  `MNEMOS_AUTH_ENABLED`.
- SQLite duplicate explicit memory IDs raise `DuplicateMemoryError`
  instead of silently succeeding.

Plumbing:
- Migrate route-level asyncpg acquires to `PoolManager.acquire()`.

## Tests

- 1087 passed, 0 failed (was 633 before v4.1).
- 9 new test files for the corpus findings (auth_failclosed,
  dag_visibility_alignment, edge_profile_unsupported_routes,
  federation_peer_ssrf, graeae_cancellation, morpheus_tenant_scoping,
  no_raw_pool_acquire, sqlite_insert_dup, webhook_ssrf_rebind).
- `tests/_fake_backend.py` Postgres-flavored `FakePoolBackedBackend`
  so consultation/audit/entity/DAG handlers see consistent test data
  through the same FakePool that backed `_pool`.

Author: perlowja

CHANGELOG.md

@@ -2,6 +2,55 @@
 
 All notable changes to MNEMOS are documented here.
 
+## [4.1.3] — 2026-04-29
+
+Corpus-review hardening release.
+
+### Fixed
+
+- Pinned webhook delivery DNS resolution from validation through HTTP connect to close DNS-rebinding SSRF.
+- Moved consultation completion and DAG live-merge webhooks into the transactional outbox path, with delivery scheduled only after commit.
+- Released GRAEAE provider concurrency slots in `finally` during cancelled fan-out.
+- Marked sessions, entities, state, and MORPHEUS HTTP routes as Postgres-only on edge profiles with explicit 503 responses.
+- Restricted MORPHEUS run telemetry reads to root/operator callers.
+- Migrated route-level asyncpg acquires to `PoolManager.acquire()`.
+- Aligned DAG read preflight with memory read visibility while keeping branch/merge writes strict-owner scoped.
+- Applied webhook SSRF URL validation to federation peers; private peer URLs require `FEDERATION_ALLOW_PRIVATE=true`.
+- Added typed `AuthSettings` and server-profile fail-closed auth defaults via `MNEMOS_AUTH_ENABLED`.
+- Made SQLite duplicate explicit memory IDs raise `DuplicateMemoryError` instead of silently succeeding.
+
+## [4.1.2] — 2026-04-29
+
+GRAEAE provider-routing fix + container-env operations standard.
+
+### Fixed
+
+- `mnemos.domain.graeae.engine._ranked_candidates` tiebreak ordering
+  added an explicit non-reasoning preference between `last_synced` and
+  `len(model_id)`. Before this fix, the `len()` fallback accidentally
+  promoted `-reasoning` SKUs (shorter names) over `-non-reasoning`
+  siblings of equal weight/version, so xAI Grok consultations came
+  back tagged with `\confidence{N}` blocks instead of clean text.
+  Provider helper `_is_reasoning_variant(model_id)` formalizes the
+  classification.
+- New regression suite at `tests/test_graeae_ranked_candidates.py`
+  covers the helper + the tiebreak ordering.
+
+### Operational
+
+- v4.x container env standard documented: every `mnemos serve`
+  container MUST mount `~/.api_keys_master.json` →
+  `/etc/mnemos/api_keys.json` (read-only) AND set
+  `MNEMOS_KEYS_PATH=/etc/mnemos/api_keys.json`. The v4.1.1 cutover
+  surfaced that without these, GRAEAE quietly falls back to
+  empty-key/no-provider state and every consultation 401s.
+- Pre-existing reasoning-variant rows in `model_registry` should be
+  marked `deprecated=true` for Grok-family providers via:
+  `UPDATE model_registry SET deprecated = true WHERE provider = 'xai'
+   AND model_id ~ '-reasoning$' AND model_id NOT LIKE '%non-reasoning'`.
+  v4.1.2 fleet rollout includes this UPDATE on PYTHIA + CERBERUS
+  before container restart.
+
 ## [4.0.0] — 2026-04-29
 
 Major refactor + multi-backend persistence + multi-worker support release.

README.md

@@ -4,7 +4,7 @@
 
 # MNEMOS + GRAEAE
 
-**MNEMOS v4.0.0 is the memory operating system for serious agentic work: a
+**MNEMOS v4.1.1 is the memory operating system for serious agentic work: a
 packaged FastAPI runtime, multi-backend persistence layer, GRAEAE reasoning bus,
 operator-audited compression stack, and CLI-first deployment surface.**
 
@@ -28,7 +28,7 @@ federate, export, import, and operate**.
   the `/v1/` REST surface are services built on top of the kernel, not retrofits
   onto a vector store.
 
-The v4.0 codebase is a coherent `mnemos/` package: `api/routes`, `core`, `db`,
+The v4.1 codebase is a coherent `mnemos/` package: `api/routes`, `core`, `db`,
 `domain`, `persistence`, `mcp`, `webhooks`, `workers`, `hooks`, `installer`,
 `tools`, and `cli`. The old top-level script sprawl is gone; operators use the
 single `mnemos` command.
@@ -61,26 +61,26 @@ single `mnemos` command.
   isolation, multi-worker smoke, and Postgres/SQLite persistence parity.
 
 MNEMOS has been in daily production use since December 2025. The current release
-line is **v4.0.0**, shipped on **2026-04-29**, and it is the first release where
+line is **v4.1.1**, shipped on **2026-04-29**, and it is the first release where
 the package layout, persistence layer, deployment profiles, CLI, single-binary
 distribution, and multi-worker coordination all match the intended production
 shape.
 
 ## Quick Install
 
 ```bash
-pip install mnemos-os==4.0.0
+pip install mnemos-os==4.1.1
 mnemos serve --profile dev
 ```
 
 ```bash
-docker pull ghcr.io/mnemos-os/mnemos:4.0.0
+docker pull ghcr.io/mnemos-os/mnemos:4.1.1
 ```
 
 For a single binary with no host Python:
 
 ```bash
-curl -L https://github.com/mnemos-os/mnemos/releases/download/v4.0.0/mnemos-linux-x86_64 -o mnemos
+curl -L https://github.com/mnemos-os/mnemos/releases/download/v4.1.1/mnemos-linux-x86_64 -o mnemos
 chmod +x mnemos
 ./mnemos install --profile edge
 ./mnemos serve --profile edge
@@ -141,7 +141,7 @@ MNEMOS was built to solve those problems in a way that reflects real platform ex
 
 Its design is informed by years of enterprise platform work, large-vendor systems thinking, open-source infrastructure experience, and current work in the AI industry, without assuming that professional users want marketing language where they really need operational clarity.
 
-**MNEMOS has been in daily production use since December 2025**, backing multiple active agentic systems simultaneously. By early 2026 the running install was holding thousands of memories and had performed thousands of compressions, each with a written quality manifest. The v3.0 release line unified that production codebase into the single-service FastAPI shape; **v4.0.0 is the current shipped GA line**, adding the coherent package layout, persistence abstraction, SQLite profiles, single-binary artifacts, and multi-worker production support. See [`CHANGELOG.md`](./CHANGELOG.md) for the release history.
+**MNEMOS has been in daily production use since December 2025**, backing multiple active agentic systems simultaneously. By early 2026 the running install was holding thousands of memories and had performed thousands of compressions, each with a written quality manifest. The v3.0 release line unified that production codebase into the single-service FastAPI shape; **v4.1.1 is the current shipped GA line**, adding the coherent package layout, persistence abstraction, SQLite profiles, single-binary artifacts, and multi-worker production support. See [`CHANGELOG.md`](./CHANGELOG.md) for the release history.
 
 For the longer story — the original catalyzing moment, the architectural decisions (and mistakes) that took MNEMOS from a single-file prototype to a unified runtime, and the scrubs, refactors, and release-gate audits that landed the public cut — see [`EVOLUTION.md`](./EVOLUTION.md). Written for future contributors as much as for future readers who want to know what they're inheriting.
 
@@ -248,7 +248,7 @@ The shared premise — that agent memory deserves first-class treatment — is t
 
 ## What works now
 
-This is the current state of the v4.0.0 release line. Features described here are implemented unless explicitly called out as forward-looking in [`ROADMAP.md`](./ROADMAP.md).
+This is the current state of the v4.1.1 release line. Features described here are implemented unless explicitly called out as forward-looking in [`ROADMAP.md`](./ROADMAP.md).
 
 The API surface is namespaced under `/v1/*`.
 
@@ -622,7 +622,7 @@ Landed with the v3.0 release line:
 
 v3.5.1 is a documentation-triage patch shipped on 2026-04-28. It bumps package/runtime version metadata to 3.5.1 and reconciles release-state docs with the v3.5.0 GA tag; it does not change product behavior from v3.5.0.
 
-### Shipped in v4.0.0
+### Shipped in v4.1.1
 
 - ✅ **Coherent package layout** — production code now lives under `mnemos/` with `api/routes`, `core`, `db`, `domain`, `persistence`, `mcp`, `webhooks`, `workers`, `hooks`, `installer`, `tools`, and `cli` subpackages.
 - ✅ **Persistence abstraction** — `PersistenceBackend` owns the contract; `PostgresBackend` uses asyncpg + pgvector + RLS + LISTEN/NOTIFY, and `SqliteBackend` uses aiosqlite + sqlite-vec + FTS5 + JSON1 + WAL.
@@ -633,9 +633,23 @@ v3.5.1 is a documentation-triage patch shipped on 2026-04-28. It bumps package/r
 - ✅ **Architectural enforcement** — seven import-linter contracts keep API, domain, db, core, persistence, MCP, and webhook boundaries honest in CI.
 - ✅ **GRAEAE mode validation** — routing modes plus `single`, `debate`, and `majority` are modeled as a `Literal`; unknown modes 422 instead of falling through.
 
-### Beyond v4.0
+### v4.1.3 known limitations
 
-Forward-looking scope is maintained in [`ROADMAP.md`](./ROADMAP.md), which lists shipped v3.x/v4.0 scope and items explicitly deferred with rationale.
+- `bulk_create_memories` now runs through the backend transaction and webhook
+  outbox surface, so it works on SQLite-backed edge profiles as well as
+  Postgres-backed server profiles.
+- The SQLite-backed `edge` profile intentionally exposes a narrower HTTP API:
+  sessions, entities, state, and MORPHEUS telemetry routes return 503 because
+  those surfaces still depend on server-profile Postgres SQL.
+- MORPHEUS run and cluster endpoints are operator-only telemetry. They require
+  root credentials because responses can include namespaces, configs, errors,
+  and memory IDs across tenants.
+- v4.1 still does not ship the separate web frontend, mobile clients, hosted
+  MNEMOS Cloud, or Rust hot-path rewrites; those remain roadmap items.
+
+### Beyond v4.1
+
+Forward-looking scope is maintained in [`ROADMAP.md`](./ROADMAP.md), which lists shipped v3.x/v4.1 scope and items explicitly deferred with rationale.
 
 Near-term not-yet-scoped candidates:
 
@@ -675,7 +689,7 @@ mnemos.core lifecycle/config/visibility
 ### Edge install (single binary)
 
 ```bash
-curl -L https://github.com/mnemos-os/mnemos/releases/download/v4.0.0/mnemos-linux-x86_64 -o mnemos
+curl -L https://github.com/mnemos-os/mnemos/releases/download/v4.1.1/mnemos-linux-x86_64 -o mnemos
 chmod +x mnemos
 ./mnemos install --profile edge
 ./mnemos serve --profile edge
@@ -684,7 +698,7 @@ chmod +x mnemos
 ### Package install
 
 ```bash
-pip install mnemos-os==4.0.0
+pip install mnemos-os==4.1.1
 mnemos install --profile dev
 mnemos serve --profile dev
 ```