Summary

Adds NamespacePolicyRule to NamespaceConfig so users can declare path-glob → namespace mappings instead of passing namespace= on every mem_index call. Closes the practical gap in enable_auto_ns, which uses the immediate parent folder name and so produces opaque namespaces like <UUID> or subagents under auto-discovered roots (~/.claude/projects/<UUID>/...).

Resolution priority

explicit param → rules (first valid match) → enable_auto_ns → default_namespace

Rules are evaluated between the explicit-argument check and the enable_auto_ns fallback. Default is [], so existing users see no behavior change until they opt in via config.json or config.d/*.json.

Design decisions

• Pattern engine: pathspec.GitIgnoreSpec, case-insensitive, matched against the absolute resolved file path with any leading / stripped. Same semantics as IndexingConfig.exclude_patterns so users learn one pattern language.
• {parent} placeholder: expands to the matched file's immediate parent folder name. Only {parent} is supported in this release; unknown placeholders are rejected at load time so typos fail loudly at startup rather than silently skipping rules at index time.
• Empty-parent fall-through: if {parent} would expand to an empty string at runtime, the rule is skipped (fall through to the next rule / auto_ns / default_namespace) with a one-time warning log per rule index.
• APPEND merge: Annotated[list[NamespacePolicyRule], APPEND] so config.d/*.json fragments can contribute rules independently. Fragments concatenate in alphabetical filename order — use numeric prefixes (10-claude.json, 20-gdrive.json) for cross-fragment precedence.
• Validation: non-empty namespace after strip, ≤ 128 chars, no ASCII control chars; non-empty path_glob with ~/ expanded at load time.

Incidental fixes to config merge machinery

This PR is the first list[BaseSettings] field in the codebase and surfaced two gaps in load_config_d that are fixed inline. Both are scoped to load time (~/.memtomem/config.json + ~/.memtomem/config.d/*.json merging) and do not touch the mutation paths (config_set CLI, PATCH /api/config, future Web UI editors).

1. _dedup_key now handles dict / BaseSettings / nested list. Previously it only normalised Path → str and let everything else fall through to raw Python hashing, which raised TypeError: unhashable type: 'dict' the first time a list-of-objects field went through APPEND dedup. The new version recursively converts BaseSettings via model_dump(mode="json") and dicts into sorted tuples, so a native model instance and the raw JSON dict that produced it share a dedup key. Existing list[str] / list[Path] / list[float] fields are untouched.
2. APPEND merge loop coerces JSON dicts to the declared item type before dedup. A non-validate_assignment BaseSettings lets setattr(section_obj, key, [<dict>]) succeed without coercing, so the merged list would contain raw dicts. The loop now introspects the field annotation, and if the item type is a BaseSettings subclass it calls item_type.model_validate(item) on each incoming dict before the dedup step. Invalid entries are logged and skipped rather than crashing the whole fragment.

Dedup semantics freeze (commit 3): _dedup_key intentionally hashes post-validator field values, so a rule written as ~/foo/** and the same rule written as the expanded absolute path /Users/X/foo/** collide after construction and dedupe to one entry. A new test (test_config_d_namespace_rules_dedup_after_home_expansion) pins this contract so a future refactor that moved ~/ expansion out of the validator — or moved the dedup key off model_dump — would fail at PR time rather than surfacing as a confused bug report from a user who split their rules across two config.d fragments.

Out of scope: the mutation path still lacks nested-BaseModel support — coerce_and_validate in config.py can't consume a list[BaseSettings] today, so config_set namespace.rules ... is rejected and a Web UI editor can't persist rules via PATCH /api/config yet. Fixing that is a separate PR (see Intentionally deferred below).

Intentionally deferred

• Web UI rules editor — blocked on the coerce_and_validate mutation-path gap above; attempting the editor before that fix would re-discover the same list[BaseSettings] problem on the mutate side.
• config_set namespace.rules CLI — same root cause as Web UI editor.
• {match:N} capture group placeholder — {parent} covers the common cases; add if evidence accrues.
• Rule priority: int field — alphabetical fragment ordering + in-fragment declaration order cover 99% of cases.
• Rule debugging CLI (mm namespace test <path>) — mm config show is enough for now.

Each is listed in the plan file's "Out of scope" block so the bounds are explicit.

Test plan

• 12 new TestNamespacePolicyRules cases (core 7 + edge 5): explicit-wins, rule-match, first-match-wins, rules-beat-auto_ns, {parent} substitution, ~/ expansion, empty-parent fall-through (with log-once assertion), case-insensitive matching, literal template, unknown-placeholder rejection, length-limit rejection.
• 3 test_config_overrides.py cases covering APPEND merge, the alphabetical-order contract, and the home-expansion dedup freeze.
• Existing TestResolveNamespace (5 tests) — regression-free.
• Existing test_every_list_field_declares_merge_strategy — passes for the new field.
• Full uv run pytest -m "not ollama" — 1650 + new 15 = 1651 passed, 46 deselected.
• uv run ruff check + uv run ruff format --check.
• uv run mypy packages/memtomem/src/memtomem/config.py packages/memtomem/src/memtomem/indexing/engine.py — no issues.

Docs in the same PR

• docs/guides/configuration.md — "Namespace rules (path-based auto-tagging)" subsection under Namespace: JSON example, full semantics, "Verifying your rules" block.
• docs/guides/user-guide.md — cross-reference from the Auto-namespace section so readers land on the rules doc when the parent folder heuristic isn't enough.