diff --git a/docs/roadmap/interfaces.md b/docs/roadmap/interfaces.md
new file mode 100644
index 0000000..57fd286
--- /dev/null
+++ b/docs/roadmap/interfaces.md
@@ -0,0 +1,197 @@
+# API and Interface Contract Audit
+
+Reviewed baseline: `01b08e76a3fecc22c4414437bfeaa07b4b7a85aa`.
+
+Status: four open findings. This is an audit record, not an implementation plan.
+No production code or repository tests were changed.
+
+## Scope and evidence
+
+Reviewed the root-owned public facade and its conversions, engine loading and
+callback boundaries, source/discovery and filesystem ownership, schema and typed
+value handling, metadata/history, safe rendering and comparison, secrets, and
+the document/YAML boundary against declaration comments, the
+[configuration contract](../config.md), and applicable architectural decisions.
+The root delegates to the synchronous engine lifecycle; parser-independent
+documents support source assembly and inheritance, and completed results retain
+detached configuration and frozen inspection state.
+
+The findings below were reproduced through the public API with a disposable Go
+program outside the repository, using labeled byte sources and `fstest.MapFS`.
+The full existing suite passed with Go 1.26.10 on Linux:
+
+```sh
+GOTOOLCHAIN=local GOTELEMETRY=off GOPROXY=off go test ./... -count=1
+```
+
+The existing tests cover many ownership, lifetime, error, and zero-value cases,
+but do not detect these cross-interface discrepancies. No benchmarks, external
+service calls, or new test files were needed. The recommendations restore
+existing contracts rather than introduce new API restrictions or guarantees.
+
+## 1. Resolution presence queries confuse explicit keys with sequence indexes
+
+Priority: P2. Status: open.
+
+Relevant code:
+
+- [`Resolution.IsExplicit`](../../internal/engine/resolution.go), lines 75–82.
+- `Resolution.InputInfo` and `resolutionInput.lookup` in
+  [resolution.go](../../internal/engine/resolution.go) and
+  [load_metadata.go](../../internal/engine/load_metadata.go).
+- `resolveSchemaPath` in [schema.go](../../internal/engine/schema.go).
+
+The public [path contract](../../path.go) distinguishes explicit mapping keys
+from sequence indexes. The [resolution contract](../config.md#recorded-application-resolution)
+defines surviving-input presence and a fixed entry-state metadata view;
+[metadata queries](../config.md#layered-metadata) preserve container shapes.
+
+`IsExplicit` validates against the schema but then looks up presence using only
+`path.String()`. In dynamic `any` state, schema validation cannot establish the
+actual container shape. Consequently these distinct queries collide:
+
+```go
+key := configkit.Keys("shape", "0")
+index := configkit.Keys("shape").Index(0)
+```
+
+Load `map[string]any` from `shape: [input]` with a resolver that queries both
+paths, without changing the configuration. Observed results:
+
+| Query | `Resolution.IsExplicit` | `Resolution.InputInfo` found | Final `Metadata.IsExplicit` |
+| --- | --- | --- | --- |
+| Explicit mapping key | true | false | false |
+| Explicit sequence index | true | true | true |
+
+The mapping-key path never existed and should not be reported as explicit.
+Applications branching on presence receive inconsistent answers depending on
+which public query they use, even during an otherwise unchanged load.
+
+Potential remediation: use the retained entry-state semantic shape when checking
+resolution presence, consistently with `InputInfo` and `Metadata.IsExplicit`.
+Review the same string-only presence lookup in `Default` while preserving its
+existing assignment and callback-lifetime rules. Add focused public coverage for
+both mapping-to-index and index-to-mapping mismatches under dynamic values,
+including contextually parsed pointers. This does not require changing the
+documented treatment of invalid schema paths.
+
+## 2. Resolution can transfer explicitness to a different semantic path
+
+Priority: P2. Status: open.
+
+Relevant code: `applyResolutionMetadata`, `applyResolutionFieldOrigins`, and
+`refreshResolutionFields` in
+[load_metadata.go](../../internal/engine/load_metadata.go), especially lines
+371–387; final shape checking in
+[`Metadata.effectiveField`](../../internal/engine/metadata.go).
+
+The [metadata contract](../config.md#layered-metadata) distinguishes key/index
+shapes and keeps derived values for absent input non-explicit. Recorded
+resolution changes origins without creating input explicitness.
+
+Reproduction:
+
+1. Load `map[string]any` from `shape: {'0': old}`.
+2. In resolution, replace `configuration["shape"]` with `[]any{"new"}` and
+   record normalization at `Keys("shape")`.
+3. Query final metadata using the two paths from finding 1.
+
+The old explicit mapping-key path correctly stops being effective, but the newly
+created sequence index reports `Explicit: true`. That sequence index was never
+supplied by input. `refreshResolutionFields` reuses the old `FieldInfo` whenever
+its textual pointer survives; it updates null state without checking whether
+the old and new containers give that pointer different semantic shapes. Final
+lookup validates only the new shape, so it cannot recover the lost distinction.
+
+This can mislead consumers inspecting whether a final value was explicitly
+configured rather than introduced by application resolution. The converse
+sequence-to-numeric-key transition has the same underlying collision.
+
+Potential remediation: retain entry/effective shape evidence across resolution
+and carry input explicitness forward only for the matching semantic path.
+Preserve existing explicitness for same-shape paths and the recorded resolution
+origin; do not clear every descendant's presence merely because an ancestor was
+normalized. Cover both shape transitions and agreement among `Lookup`,
+`IsExplicit`, and explanation. Fixing finding 1's callback query alone does not
+repair this publication path.
+
+## 3. Metadata history mixes numeric-key and sequence-index definitions
+
+Priority: P2. Status: open.
+
+Relevant code:
+
+- [`Metadata.History`](../../internal/engine/metadata.go), lines 56–65.
+- `metadataFromDecoded`, `applyOverlayMetadata`, and history event construction
+  in [load_metadata.go](../../internal/engine/load_metadata.go).
+- Shape-aware replay in [`Inspection.Explain`](../../internal/engine/explain.go).
+
+The [metadata contract](../config.md#layered-metadata) promises semantic
+key/index distinctions for historical as well as effective queries. The
+historical-path index retains both shapes, but `History` uses that index only to
+check that the query was ever known. It then retrieves the entire history bucket
+by textual pointer, without selecting the query's shape.
+
+Reproduction: load `shape: {'0': old}` into `map[string]any`, then apply a CLI
+override at `Keys("shape")` with `[]any{"new"}`. Both explicit shapes were
+observed and share `/shape/0`.
+
+`History(Keys("shape").Index(0))` incorrectly includes the original base
+definition of the mapping key, and identifies that base origin as the previous
+origin of the CLI operation. The old input defined a key, not an index. In the
+same successful result, `Explain` for the explicit index correctly shows only
+its CLI introduction. `History` for the key and index returns identical records
+despite their different lifetimes.
+
+Consequences include false source attribution and disagreement between the two
+public history interfaces when consumers inspect removed or introduced paths.
+
+Potential remediation: retain enough container-shape evidence with stored
+history events, or select relevant events from existing shape-aware replay
+evidence. Filter exact definitions and previous origins by the queried semantic
+shape while retaining genuine ancestor replacement/removal events. Filtering
+only on the current `Change.Path` is insufficient where construction already
+replaced concrete segment kinds with parsed tokens. Preserve the intentional
+one-pointer-per-entry `HistoricalPaths` enumeration and contextual behavior of
+`ParsePath`; neither public contract needs redesign. Cover both direction changes
+and compare metadata history with explanation.
+
+## 4. Diagnostic canonicalization breaks composition claimant ordering
+
+Priority: P3. Status: open.
+
+Relevant code: `compositionConflictDiagnostic` in
+[compose.go](../../internal/engine/compose.go) and `canonicalDiagnostics` in
+[diagnostic.go](../../internal/engine/diagnostic.go), particularly lines 88–92.
+The latter runs both while constructing `LoadError` and in `WriteDiagnostics`.
+
+The [strict-composition contract](../config.md#strict-document-composition)
+specifies the first claim in scoped assembly order as primary and every remaining
+claim in the same order as related locations.
+
+A native-filesystem fixture demonstrates the discrepancy:
+
+```text
+main.yml imports a.yml, b.yml, c.yml
+a.yml imports z.yml
+b.yml, c.yml, and z.yml each define x
+```
+
+Acquisition assigns main=source-1, a=source-2, z=source-3, b=source-4,
+c=source-5. Dependency-ready assembly considers the three conflicting claims
+in b, c, z order. The composer correctly chooses b as primary and constructs
+related locations c, z. Generic canonicalization sorts those related locations
+by acquisition ID, publishing z, c instead. The reproduced public diagnostic
+has primary source-4 and related source-3, source-5.
+
+All claimants remain present and the conflict still fails loading, so this is
+lower impact than the presence/history defects. It nevertheless violates the
+documented reporting order and gives consumers two different ordering rules
+within one conflict report.
+
+Potential remediation: preserve the selected semantic ordering for composition
+claimant lists through accumulation and rendering, while retaining canonical
+ordering of diagnostic records and the established sorting of related locations
+for diagnostics without a semantic-order contract. Add one public failure-path
+case whose acquisition and claimant orders differ and verify both structured
+diagnostics and rendered output. No scheduler or precedence change is needed.