diff --git a/docs/roadmap/internal-boundaries.md b/docs/roadmap/internal-boundaries.md
new file mode 100644
index 0000000..149e7d1
--- /dev/null
+++ b/docs/roadmap/internal-boundaries.md
@@ -0,0 +1,191 @@
+# Import Discovery And Facade Boundary Simplification
+
+Status: target defined; staged implementation planning pending.
+Baseline: `80cf23e7102fca2c5d7c7621e75f930a1ef1ecad`.
+
+This roadmap brings together the two refactors deferred until after the audit
+remediation sprint: glob entry preparation (**C03**) and facade-to-engine
+transport simplification (**F07**). Their original motivation is retained in the
+[glob](future.md#1-separate-glob-entry-preparation-from-traversal) and
+[transport](future.md#2-simplify-redundant-facade-to-engine-transport) future-work
+entries and the [consolidated audit](audit.md).
+
+The goal is clearer responsibility and ownership with fewer redundant internal
+representations. Both changes preserve the public API and configuration model.
+They are independent work areas within one maintenance effort, not a new loading
+workflow. A subsequent implementation plan will own concrete edits, stages,
+validation commands, and execution progress.
+
+## Baseline And Targeted End State
+
+The preceding sprint repaired directory-failure classification and shared link
+dependency evaluation, clarified acquisition handoffs, and removed unused
+placement bridges. Those changes provide the baseline; this work does not repeat
+or redesign them.
+
+| Area | Current baseline | Target |
+| --- | --- | --- |
+| Glob discovery | `selectGlobImportCandidates` interleaves traversal and set operations with substantial per-entry preparation. | The selector coordinates discovery; a cohesive private operation prepares a wildcard-discovered entry. |
+| Facade transport | Import defaults pass through root conversion, an exported engine transport record, an equivalent private engine record, and repeated snapshots. Placement construction also uses setter-style bridges. | Root-owned public records convert explicitly to one engine defaults representation, with straightforward engine option/input construction and identifiable ownership copies. |
+
+The [architecture policy](../policy/architecture.md), [configuration
+contract](../config.md), and public declarations remain authoritative. [ADR
+0006](../adr/0006-public-facade.md) continues to govern the facade boundary. These
+refactors do not require a new architectural decision or changes to its accepted
+decision. The [component overview](../internal/overview.md) and [import
+internals](../internal/imports.md) own implemented structure.
+
+## Glob Entry Preparation
+
+### Target Responsibility Split
+
+In [glob.go](../../internal/engine/glob.go), the selector continues to own pattern
+validation, component advancement, literal-segment handling, directory
+enumeration and its accounting, candidate union and sorting, final target
+validation, exclusions, and requiredness.
+
+A private glob-specific preparation operation owns the patterned branch's work
+for one directory entry: matching and hidden-entry filtering, lexical/access path
+construction, containment checks, target inspection and discovered-link policy,
+intermediate-directory eligibility, host/native access-path selection, labels,
+and candidate construction. Its result distinguishes a selected candidate, an
+ordinary skip, and a diagnostic. It receives the facts needed for that entry
+rather than mutating the selector's candidate set or surrounding traversal state.
+
+Use the existing filesystem access and link-resolution helpers. Directory
+selection already has `prepareImportDirectoryEntry`; it remains a separate owner
+because its recursion, matching, absence handling, and traversal state differ.
+Share a smaller operation only where it actually expresses the same semantics
+and reduces duplication. No common visitor or discovery strategy is needed for
+this extraction.
+
+### Behavior To Preserve
+
+Preserve the current distinctions and processing order documented by
+[explicit imports](../config.md#explicit-imports) and exercised by the selection
+tests:
+
+- Literal components remain explicit selections; discovery filters apply at
+  their existing wildcard boundaries. Moving matching or hidden filtering after
+  filesystem inspection would change which errors are observable.
+- Keep lexical candidate identity separate from resolved access identity.
+  Native source identity and relative imports retain namespace spelling; host
+  access retains its existing canonicalization and containment behavior.
+- Keep discovered-link follow rules, intermediate-directory eligibility,
+  missing-target handling, and structured diagnostic sites/classifications.
+  Do not make glob and directory rules identical as a side effect of extraction.
+- Preserve deterministic candidate and error order, union/deduplication,
+  directory-entry accounting, and the existing point of source admission.
+  Final selected-file validation remains before exclusions where it is today;
+  exclusions do not hide discovery or resource errors.
+- Retain the preceding sprint's distinction between missing-on-open directories
+  and failures after opening, along with operation-local completed-link reuse.
+
+The intended benefit is that a maintainer can understand traversal separately
+from entry eligibility and identity preparation. This is a readability change;
+it does not select a different search algorithm or claim a speedup.
+
+## Facade-To-Engine Transport
+
+### Target Representations
+
+The root keeps its public `ImportDefaults`, `PlacementMode`, `Options`, `Input`,
+and `LayeredOptions` types. Explicit conversions remain appropriate at this
+boundary; consumers must not acquire engine type identity, fields, or methods.
+
+Inside the engine, consolidate the equivalent `ImportDefaults` transport and
+private `importDefaults` records into one presence-aware input-defaults
+representation. Carry that representation through the ordinary and layered
+options, normalized request, and acquisition policy instead of translating the
+same five fields between equivalent engine records.
+
+Keep `resolvedImportPolicy` distinct: it represents concrete per-declaration
+choices, including selector-specific rules, rather than another copy of optional
+load-wide input. Likewise, registration compilation and the normalized loading
+request have real responsibilities and are not replacement targets.
+
+Make placement and defaults straightforward to construct on the existing engine
+option/input records. Remove decorator-style bridges made unnecessary by that
+construction surface. Exporting fields or types within `internal/engine` where
+needed by the root converter does not make them part of ConfigKit's supported
+public API. Keep explicit enum conversion and preserve unknown values until the
+existing option validation can report them; conversion must not silently choose
+a default or move validation into facade setters.
+
+### Ownership And Copying
+
+Make the actual mutable ownership boundaries identifiable. Public load calls
+continue to snapshot pointed-to default booleans and the pattern slice before
+source acquisition, as specified by the [public
+declaration](../../lifecycle.go). Preserve omitted, explicit false, nil pattern,
+and non-nil empty pattern as distinct inputs where they are distinct today.
+
+Review the entire handoff, not just the duplicate type definitions: root
+conversion in [conversion.go](../../conversion.go), ordinary construction in
+[lifecycle.go](../../lifecycle.go), layered construction in
+[layered.go](../../layered.go), engine options and normalization, acquisition
+policy construction, and declaration resolution in
+[import_spec.go](../../internal/engine/import_spec.go).
+
+Remove snapshots that merely recopy already-owned read-only defaults as they
+cross those private representations. Keep copies needed when establishing
+independent mutable ownership, including any independently owned resolved
+policy. The target is explainable ownership and less redundant work, not an exact
+copy count or removal of copying throughout the library. Shared defaults within
+one load can be read-only by implementation discipline; this creates no new
+caller immutability or concurrency requirement.
+
+Preserve ordinary and layered option preflight before source reads and application
+hooks, independent declaration overrides, transitive/underlay propagation,
+selected-overlay behavior, destination normalization, and invalid-enum
+diagnostics. Keep callback, source, path, secret, and returned-record conversions
+whose ownership or domain meaning is not redundant transport.
+
+The intended benefit is fewer concepts, field mappings, and construction paths
+to maintain. Reduced copying is useful but does not establish an end-to-end
+performance guarantee.
+
+## Implementation And Documentation Work
+
+The later plan should identify the glob extraction boundary and map the defaults
+and placement ownership handoffs against current callers before changing them.
+The two refactors can be implemented and reviewed independently; neither needs
+the other to land. Retire replaced private representations and helpers together
+with their callers rather than adding a second production path.
+
+Update internal documentation when the resulting boundaries land. Public
+declaration wording needs changes only if required to keep its existing
+ownership description accurate; this effort does not select new public behavior.
+Replace the corresponding C03/F07 candidate detail in `future.md` with a pointer
+to the active work when the planning records are reorganized, avoiding parallel
+target or progress inventories. Leave the unrelated future candidates alone.
+
+## Validation And Completion
+
+Follow the [testing policy](../policy/testing.md) and [development
+guide](../development.md#baseline-validation). Reuse the existing glob selection,
+directory/link failure, import-default, placement, public facade, and dependency
+boundary coverage. Add a focused behavioral case only where a concrete
+extraction or ownership risk lacks protection.
+
+For discovery, verify literal versus wildcard filters, host/native identity,
+contained and escaping links, optional absence versus read/close failures,
+ordering, exclusions, regular-file checks, and existing resource accounting. For
+transport, verify nil/empty and explicit-false distinctions, mutation of caller
+default storage after the load snapshot through a supported callback, invalid
+options before input consumption, both loading APIs, and transitive/underlay
+defaults. Existing conversion-focused tests may change with private construction;
+preserve the useful ownership behavior rather than fossilizing a helper's shape.
+
+Completion requires an easier-to-follow selector, one engine input-defaults
+representation, fewer redundant mapping/construction sites, justified remaining
+copies, and preserved public behavior under the relevant checks. Review source
+and ownership flow as well as passing tests; moving complexity into a helper or
+renaming a duplicate record alone is insufficient.
+
+No benchmark campaign, complexity score, allocation threshold, exact filesystem
+call trace, new resource limit, or additional release gate is required. This work
+adds no dependencies, caches, background activity, or new consumer controls, and
+does not reopen other audit findings or deferred performance experiments. Routine
+helper signatures, names, and file placement belong to the implementation plan;
+no significant product or architectural decision remains open for this scope.