Summary

PR #253 introduced NamespacePolicyRule as memtomem's first list[BaseSettings] config field. The load path was wired to handle it correctly (via _list_item_type + model_validate in load_config_d and _dedup_key), but the mutation path was left asymmetric — coerce_and_validate, PATCH /api/config, mm config set, and the init wizard all silently passed raw dicts through. Downstream code (indexing/engine.py:121) accesses rule.path_glob on each entry and would AttributeError on a dict.

Adjacent to PR #256 (save-path drop-default); together #253 → #256 → this PR complete the config-layer round-trip.

Root cause

PR #253 added the new list[BaseSettings] shape but only fixed the read side of the config pipeline. The write side — the three canonical mutation entry points plus the init wizard — kept validating only primitive lists. A user PATCH-ing namespace.rules or running mm config set namespace.rules '[...]' saw no validation error, but the runtime config held dicts where model instances were expected.

What this does

1. coerce_and_validate handles list[BaseSettings]

New branch inside the existing expected_type is list path, gated by:

if isinstance(item_type, type) and issubclass(item_type, BaseSettings):

The isinstance(item_type, type) guard is deliberate — it protects against Union, generic aliases, or other non-class item types that would TypeError inside issubclass. Existing primitive-list callers (search.rrf_weights, importance.weights) take the else path unchanged; all their existing tests still pass.

Inside the branch: JSON strings are parsed (CLI path), lists of dicts are model_validate-ed (Web UI path), already-instance entries pass through, scalars/malformed shapes raise ValueError that propagates as a rejected entry in PATCH or a non-zero CLI exit.

2. MUTABLE_FIELDS["namespace"] + FIELD_CONSTRAINTS["namespace.rules"]

Registers the field at both gates so PATCH /api/config no longer reports namespace.rules: read-only field and mm config set no longer prints not a mutable field. Frontend impact check: grep FIELD_CONSTRAINTS packages/memtomem/src/memtomem/web/static/ → zero hits. UI rendering is unaffected.

3. Unified _json_default JSON serializer

save_config_overrides previously used json.dumps(..., default=str), which would emit BaseSettings.__repr__ for any nested model — a latent round-trip bug. Fixed with _json_default that returns model_dump(mode="json") for BaseSettings and str() for Path. Scope audit via grep "default=str\|default=_json_default" packages/memtomem/src/ confirmed:

• config.py save_config_overrides → fixed (uses _json_default)
• cli/init_cmd.py:706 wizard write → also fixed (imports + uses _json_default) for consistency across all config-write paths
• 7 other default=str call sites (chunking/structured.py, cli/watchdog_cmd.py, cli/config_cmd.py show, server/scheduler.py, server/health_store.py, server/webhooks.py, server/tools/consolidation.py) all serialize non-config payloads (scheduler state, webhook bodies, health snapshots, etc.) — unrelated to BaseSettings config models, intentionally left alone.

Though namespace.rules is currently the only list[BaseSettings] field in play, unifying the helper prevents a future nested model from silently re-introducing the same latent bug.

4. Migration safety for pre-fix installations

Any config.json written by a pre-fix wizard may contain repr-string leftovers for namespace.rules. On load:

• The entry hits coerce_and_validate via FIELD_CONSTRAINTS["namespace.rules"].
• Malformed shape (scalar string, list-of-strings) → ValueError.
• load_config_overrides catches the ValueError, logs a warning, skips the key.
• The field falls back to its default []. No crash, no data loss beyond the already-corrupt entry.

Test: test_legacy_repr_string_in_config_handled_gracefully covers both scalar-value and list-of-strings cases.

Testing

• 1694 tests pass (+14 new), ruff + format + mypy clean.
• New tests in test_cli.py::TestCoerceAndValidate:
  • test_namespace_rules_from_json_string — CLI path
  • test_namespace_rules_from_list_of_dicts_pr253_regression — Web PATCH path; docstring pins the engine.py:121 failure mode that originally motivated this fix
  • test_namespace_rules_passthrough_for_model_instances
  • test_namespace_rules_empty_list
  • test_namespace_rules_rejects_malformed_json
  • test_namespace_rules_rejects_non_list_json
  • test_namespace_rules_rejects_scalar_entry
  • test_namespace_rules_propagates_model_validation_error
• New tests in test_cli.py::TestSaveConfigOverrides:
  • test_legacy_repr_string_in_config_handled_gracefully — migration safety
  • test_namespace_rules_round_trip — save → load returns validated instances (not dicts)
• New tests in test_cli.py::TestConfigCLI:
  • test_config_set_namespace_rules_json — end-to-end CLI persistence + reload
  • test_config_set_namespace_rules_rejects_malformed
• New tests in test_web_routes_extended.py::TestConfigPatch:
  • test_patch_namespace_rules — PATCH apply + runtime type check
  • test_patch_namespace_rules_validation_error — bad rule lands in rejected
• Manual smoke (4-step): mm config set namespace.rules '[...]' → reload produces NamespacePolicyRule instances → _FRESH_PRESERVE_KEYS contains namespace.rules (PR cli(init): add --fresh to drop wizard-untouched config leftovers #255 preserve list) → subsequent save_config_overrides keeps the rules (non-default, so not dropped by PR fix(config): drop default-valued fields on save to prevent silent leftover pins #256 logic).

Unblocks

• Web UI namespace rules editor (designed in project_list_basesettings_config_gap.md)
• mm init wizard rule presets
• Any future list[BaseSettings] config field

Test plan

• uv run pytest -m "not ollama" — 1694 passed
• uv run ruff check + ruff format --check — clean
• uv run mypy packages/memtomem/src — 0 issues
• Manual smoke: mm config set namespace.rules '[...]' round-trip, reload produces model instances
• Manual smoke: _FRESH_PRESERVE_KEYS contains namespace.rules (interop with PR cli(init): add --fresh to drop wizard-untouched config leftovers #255)