Problem

_revert_to_stored in packages/memtomem/src/memtomem/server/tools/status_config.py:254,256,266 writes to app.embedder, app.search_pipeline, app.index_engine:

def _revert_to_stored(app: AppContext) -> str:
    ...
    new_embedder = create_embedder(config.embedding)
    app.embedder = new_embedder                  # line 254
    app.search_pipeline = SearchPipeline(...)    # line 256
    ...
    app.index_engine = IndexEngine(...)          # line 266
    ...

After #399 Phase 1 (7855042), these three names on AppContext are read-only properties that proxy to self._components.<name>. Assigning to a property without a setter raises AttributeError: property 'embedder' of 'AppContext' object has no setter at runtime.

mypy flags it:

status_config.py:254: error: Property "embedder" defined in "AppContext" is read-only  [misc]
status_config.py:256: error: Property "search_pipeline" defined in "AppContext" is read-only  [misc]
status_config.py:266: error: Property "index_engine" defined in "AppContext" is read-only  [misc]

Because mypy is advisory in this repo (CLAUDE.md) and there is no test that calls mem_embedding_reset(mode="revert_to_stored") end-to-end (grep: only _maybe_offer_embedding_reset in tests/test_init_cmd.py, which hits a different code path), the regression shipped with Phase 1.

Impact

mem_embedding_reset(mode="revert_to_stored") is a recovery path — uncommon, but the whole point is that it stays callable when the DB is in degraded mode (#349). An AttributeError in the recovery tool defeats that. The user's fallback is re-starting the server after manually editing config, which is exactly the friction mem_embedding_reset exists to remove.

Fix sketch

Components is a mutable @dataclass (server/component_factory.py:30-44), so the natural fix is to mutate the inner container:

comp = app._components
assert comp is not None  # handler already awaited ensure_initialized
comp.embedder = new_embedder
comp.search_pipeline = SearchPipeline(...)
comp.index_engine = IndexEngine(...)

After this, app.embedder (the property) reads back the new value — no divergence.

Alternative: add explicit mutator methods on AppContext (replace_embedder / replace_search_pipeline / replace_index_engine) so the _components access stays behind the class's public surface. Probably overkill for a recovery-only path.

Test gap to close in the same fix

Add an integration test that calls mem_embedding_reset(mode="revert_to_stored") on a fixture in degraded mode and asserts:

1. The tool returns the success message without raising.
2. app.embedder / app.search_pipeline / app.index_engine reflect the stored (not configured) embedder dimension after the call.
3. A follow-up mem_search (or any vector-dependent tool) succeeds — demonstrating the three slots were actually swapped, not just embedder.

Context

• Introduced by: Lazy DB init: defer ~/.memtomem/memtomem.db creation until first tool call #399 Phase 1 (7855042 feat(server): add AppContext.ensure_initialized plumbing) — the property conversion of storage / embedder / index_engine / search_pipeline. Those properties exist so AppContext can defer component creation until first tool call; the read-only shape is deliberate but this one write-path wasn't migrated.
• Surfaced by: Lazy DB init: defer ~/.memtomem/memtomem.db creation until first tool call #399 Phase 2 handler migration work — mypy packages/memtomem/src/memtomem/server output during the PR's lint pass.
• Not a Phase 2 fix because: Phase 2 is handler migration only. Bundling this would mix concerns (feedback_one_change_per_pr).