ConfigKit/docs/config.md

# Configuration Input Contract

This document owns ConfigKit's currently implemented source-input, YAML syntax,
schema, typed-decoding, loading, metadata, redaction, rendering, and ownership
contracts. The parser, owned document representation, and decoder remain private
behind the public loading lifecycle.

## Sources

ConfigKit exposes opaque sources through these constructors:

| Constructor | Current contract |
| --- | --- |
| `File(path)` | Requires a nonempty path, opens a regular file for each consumption, and closes only the library-owned file handle. Missing, inaccessible, unreadable, directory, and other nonregular inputs fail. |
| `NewFilesystem(label, filesystem).File(name)` | Borrows a caller-owned `fs.FS` under a safe nonempty display label. `name` must be a non-root slash-relative `fs.ValidPath`. Each load reopens the selected regular file; copies of the `Filesystem` preserve namespace identity. Separate constructors create separate namespaces even for the same label or filesystem. |
| `Bytes(label, data)` | Requires a nonempty, non-secret diagnostic label and snapshots bytes during construction. Later caller mutation cannot affect the source. |
| `Reader(label, input)` | Requires a nonempty, non-secret label and non-nil reader. The reader is claimed once across all copies of the source, consumed but never closed, and cannot be reused. |
| `Mapping(label, values)` | Requires a safe nonempty label and a non-nil root `map[string]any`. Snapshots supported built-in values into an owned mapping document at construction; `{}` is valid. Constructor conversion failures become safe deferred diagnostics. |
| `NoSource()` | Selects defaults-only loading, creates no source record, and marks no paths explicit. |

The zero `Source` is invalid. Constructor argument failures are retained inside the
opaque source and become safe structured diagnostics when the source is consumed.
File and filesystem sources can observe changes between consumptions; byte and
mapping sources cannot. Mapping accepts nil, built-in string, boolean, signed or
unsigned integer, finite float, `[]any`, and `map[string]any` values recursively.
Typed nil maps and slices represent explicit null; empty collections remain empty.
Keys are sorted for stable provenance, integer decimal spelling stays exact, and
cycles or unsupported types are rejected without exposing values. Mapping input
uses the same directive and version rules as YAML, not CLI override restrictions.

Each source is limited to 8 MiB. A load using imports or an active overlay is limited
to 64 MiB across all accepted sources, 1,024 accepted sources, 64 active import
edges, and 1,000,000 input nodes in aggregate. Glob expansion separately permits
1,000,000 visited directory entries, including nonmatches. Reads request at most the
remaining allowance plus one byte so excess data is rejected rather than silently
truncated. Data returned together with a read error is not accepted as partial input.
For mapping sources, key and string bytes use the same accounting as programmatic
overrides, and the converted tree uses the same node and depth bounds.

## Single-source loading

`Load[T](source, options...)` accepts zero or one `Options[T]`. The configuration
root `T` must satisfy the application schema contract below. With no options it
starts from the Go zero value. A `Definition[T].Defaults` factory, when supplied,
runs exactly once after successful source parsing, schema construction, optional
schema-version checking, and selected-overlay loading. Its result is detached before
input is applied. Resolver-provided mutable values are detached before secret lookup
and validation, and a final detached copy prevents equal-valued aliases introduced
by validators from escaping into the returned configuration. The publication copy
preserves already validated secret state without rereading the environment.

The current lifecycle is:

```text
validate options → parse the selected source → extract source controls and
  traverse/strictly compose enabled imports
  → build schema → check the optional application version → apply a selected overlay
  → expand sibling inheritance
  → create and clone defaults → strictly decode → apply explicit CLI overrides
  → recorded application resolution
  → resolve typed secret references → semantic validation → classify secrets
  → snapshot configuration and metadata
```

`File`, `Filesystem.File`, `Bytes`, `Reader`, and `Mapping` each contribute the first base source with deterministic
ID `source-1`; enabled imports continue in depth-first order as `source-2` onward.
`NoSource` skips parsing, supplies an absent version to a configured version checker,
contributes no source, and continues through defaults, decode, and projection. A
missing file remains an error; it is never interpreted as no input.

`Options` accepts the application definition, ordinary-source `Destination` and
`PlacementMode`, explicit imports, one selected overlay, filesystem
context, explicit CLI overrides, and shared source-version controls. The definition
may register sequence identities through `Definition.SequenceIdentities`. Its
zero-valued placement mode places the source mapping
at the destination; `sequence-item` explicitly wraps the mapping as one item at a
non-root destination. Unknown modes, invalid destinations, and invalid or
conflicting sequence registrations fail before source acquisition. Overlay and
CLI coordinates remain rooted at the application root. Ordinary `Load` recognizes
source-root collection declarations whether or not imports are enabled; an
`imports` declaration still requires imports to be enabled. `NoSource` may enable
imports, but it has no main declarations to traverse. It accepts an explicit
filesystem context: a host context requires an absolute existing `Root` and may
omit `BaseDir` to use that root as its internal import base; a native context
requires its namespace, `BaseDir`, and `Root`. Supplied context is validated even
when imports are disabled. An active ordinary overlay requires this explicit
context. `NoSource` accepts overrides and remains absent rather than contributing
an empty document.

Every source, options, schema, version, defaults, decode, resolution, or validation
failure returns a nil result and a `*LoadError`. Successful results contain caller-owned `Config`, detached
`Metadata`, a defensive source list, and defensive nonfatal diagnostics. Each source
record's `Kind` identifies file, filesystem, bytes, reader, mapping, or CLI acquisition
independently of its precedence `Layer`. Host file paths remain canonical;
filesystem paths are slash-relative within their namespace; synthetic inputs have
no path. Mutating
the returned configuration or inspection records cannot alter ConfigKit's private
snapshot. ConfigKit performs no logging. `Result.WriteEffective` exposes only the
frozen redacted projection.

### Completion and resource lifetime

`Load` is synchronous and complete on return. Success means input assembly,
defaulting, typed decoding, configured application resolution, secret lookup,
validation, and configuration/metadata snapshotting have completed. A reported
load failure returns diagnostics and a nil result, as described above.

Before returning on success or failure, ConfigKit closes the filesystem handles
it opened for loading, including files and directory/root handles. Caller-owned
filesystems are borrowed, never closed. The returned
configuration and inspection metadata require no close operation and remain
usable independently of the source files. A supplied `Reader` remains
caller-owned and is never closed by ConfigKit.

ConfigKit performs no background work after returning: it does not watch inputs,
refresh secrets, or mutate published results. Inspection and rendering run only
when explicitly called and use the retained snapshot. Applications normally load
once near startup, but may explicitly load again using sources that satisfy their
documented reuse rules. Additional loads produce independently owned results.

Completion is not a transaction over external state. Files and environment
variables are observed during processing rather than as one simultaneous external
snapshot. ConfigKit does not roll back application callback effects or manage work
or resources created by application callbacks. The explicit secrets-directory
bootstrap remains a separate operation invoked by the application.

## Layered loading

`LoadLayered[T](main, options...)` accepts zero or one `LayeredOptions[T]` value.
Its `main` is an `Input` containing a source, import opt-in, and optional
filesystem context. A zero source is invalid; `Input{Source: NoSource()}` omits
the main document. Each `Underlay` contains inputs that compose strictly as
one tier, including each input's enabled imports. Empty groups contribute
nothing. Underlay groups are listed from lowest to highest priority. Repeating
a file within one group remains an import/repetition error; the same reusable
source may be selected again in another tier.

ConfigKit acquires the main input first and constructs the schema. It then
acquires and strictly composes underlay groups from lowest to highest priority,
passing each group's accumulated lower collection policy into the next group.
Each group is applied once as a precedence tier. The main tier is then composed
with the accumulated underlay policy, and the optional application version is
checked against that composed main tier only. Underlay roots cannot claim that
version path (`underlay.version`); imported fragments retain `import.version`
protection. ConfigKit then applies the selected overlay and resolves inheritance.
Scalar, sequence, and null definitions replace lower values between tiers unless
a collection policy defines the sequence operation; mappings combine recursively.
After successful defaults and typed decoding, CLI overrides are applied to that
typed configuration before application resolution. Syntax, strict-composition,
inheritance, and decoding errors cannot be masked by a CLI override. Resolution,
secret lookup, validation, and publication run once on the resulting values.

An active layered overlay requires `LayeredOptions.OverlayFilesystem` explicitly,
independent of the main source's namespace. Its name and root-relative lookup
follow the selected-overlay rules below; inactive overlays ignore the context.
The same path in a different filesystem namespace is not a repeated file.
Ordinary `Load` infers overlay context from sources that provide one; with
`NoSource`, it requires the explicit `Options.Filesystem` context.

Source IDs follow acquisition order—main and its depth-first imports, then each
underlay tier's inputs and depth-first imports, then overlay and CLI—even though
history follows precedence order. Default main-version checking happens after
underlay composition because the main tier inherits the accumulated collection
policy. Enabled per-source version checks retain acquisition order and source
local headers. Source `Layer` describes precedence while
`Kind` describes acquisition. Underlay records use `LayerUnderlay` and a
one-based `UnderlayIndex` giving their original tier position; all other records
have index zero. Explicit empty groups and `NoSource` consume no source ID.
Byte, source, node, and glob-entry budgets are shared across the load, while
the existing per-source and import-depth bounds still apply.

When a source mapping is explicitly placed as an item, its effective destination
is the sequence container. After assembly assigns an index, nested source
destinations are projected through that item. Physical source locations and
resource contexts continue to identify the declaring source.

#### Input destinations

`Input.Destination` places the source's application mapping beneath a path
relative to the configuration root. Its zero or empty value means root, so
existing inputs keep their behavior. Main inputs and underlay inputs use the
same destination model. Destination segments are literal mapping keys: dots,
slashes, and numeric-looking text are not split or interpreted as sequence
indexes. Explicit index segments and paths that introduce the reserved
`configkit` or mapping-valued `extends` structure are invalid input options and
produce a structured `LoadError` before source acquisition.

Source-root `configkit.imports` directives are processed before placement. An
import without an explicit destination inherits the effective destination of
its importing document, including through nested imports. An import may instead
declare a relative destination as a sequence of literal mapping keys; its keys
append to the importing document's effective destination. Omission and `[]`
add no keys. Nested imports continue relative to their immediate importing
document, each declaration is independent of its siblings, and all matches of
one glob receive the same resulting destination. This logical placement does
not change how the physical import path resolves from its declaring file.

A valid destination on `NoSource()` contributes nothing; an actual empty
mapping materializes its destination containers. Once placed, contributions
follow the same strict composition within a tier and recursive precedence
behavior described above. A selected overlay and CLI overrides remain rooted at
the application root and are not moved beneath the main input's destination.

`SourceInfo.Destination` reports each acquired source's effective mapping or
sequence-container path, including imports. `PlacementMode` reports how that
source was placed; mapping is the zero value, and CLI/overlay sources report
mapping unless item placement was applied. `DestinationUnresolved` is true only
when an enclosing item index needed to name the source's destination has not yet
been assigned. In that case `Destination` reports the nearest resolved enclosing
collection container and omits the unresolved local suffix. A source's own item
placement still has an exact destination at its sequence container. Completed
loads project nested source destinations through their owning items. Callback and
failed-load snapshots preserve the evidence available at that point; a fallback
snapshot remains unresolved, while a later exact snapshot is false. Returned
source records and paths are detached. Placement does not change source IDs,
acquisition identity, physical source locations, or passive resource contexts.
Introduced ancestor mappings use the contributing document's root location as a
source fallback; authored content retains its original source coordinates.

The maintained [layered consumer example](consumers/README.md#native-filesystem-and-layered-input)
shows a main host file, two underlay tiers, embedded defaults, a mapping source,
and explicit resource access without a second configuration pass.
The accepted rationale for input placement and its relative-import extension
is recorded in [ADR 0011](adr/0011-input-destinations.md) and
[ADR 0012](adr/0012-relative-import-destinations.md).

### Sibling inheritance

Any input mapping may declare `extends: name` to inherit from the mapping named
`name` among its immediate siblings. Names are exact keys, including the empty
string; lookup does not search ancestor or unrelated scopes. A mapping inside a
sequence can also participate in record inheritance when that sequence has an
effective identity declaration from `configkit.sequences` or an application
registration. In that case, `extends` is a mapping containing exactly the
sequence's identity fields, with scalar values; lookup is limited to that final
sequence. The document root and a sequence-element mapping without an effective
identity declaration have no sibling namespace for non-null string
relationships, though mappings nested inside them can have one. An `extends`
value of null cancels a relationship without looking up a parent. Other value
types, incomplete or extra reference fields, missing parents, and cycles fail
the load with value-free diagnostics.

Within each mapping sibling scope or identified sequence, direct definitions
resolve by dependency readiness. A parent is resolved before a child; among
currently eligible definitions, the lexically smallest canonical semantic
pointer is selected, including newly eligible definitions as soon as their
parent completes. This ordering governs inheritance operations and their
history, controls, and transfers. It does not reorder mappings or sequence
records in the resulting configuration.

`extends` is reserved structural input across ordinary and layered loads, every
source kind, imports, selected overlays, and explicit CLI overrides. It is
consumed before typed decoding and need not appear in the application's Go
types. Applications previously using `extends` as an ordinary YAML field must
rename that key and its field. No collection registration or keyword setup is
required for mapping-sibling relationships; identified-sequence relationships
require effective identity fields for that collection.

ConfigKit completes strict composition within tiers and precedence assembly
between tiers before resolving relationships. Only records that survive that
assembly participate in identified-sequence lookup: an upserted parent supplies
its final record, a replaced parent is unavailable, and a higher replacement
without `extends` supersedes a lower relationship. An omitted `extends` retains
a lower-priority mapping-sibling relationship; an identified record's effective
directive depends on the record that survives assembly. `extends: null` cancels
a relationship while preserving the child's own settings. Superseded references
are historical evidence, not active relationships to validate. Each child
recursively overlays its own settings over its resolved parent: explicit zero,
false, empty string, and
null values replace inherited values, while mappings combine recursively. For
identified records, the child's identity fields and their origins remain its
own; parent identity fields do not contribute to the payload.
Sequences at collection destinations use the shared collection operation with
the parent as lower input and the child as higher input. The destination's
explicit mode applies; otherwise a copied identity selects upsert and an
unkeyed destination replaces. A parent's explicit mode does not move to a
different destination. Sequences outside collection destinations remain atomic
replacements.
All effective relationships resolve, including
entries the application may never select for execution. Partial parent
definitions undergo normal typed decoding and application-owned validation,
not a special structural exemption.

Enclosing relationships resolve first. Nested instructions carried by a parent
are then resolved separately at each destination, so a child's nested sibling
override can change an inherited nested relationship. Typed defaults and
application-derived values do not start another inheritance pass. Inherited
input counts as explicit presence for `Metadata.IsExplicit`,
`Resolution.IsExplicit`, and absence-based resolver defaults, even for false,
zero, empty, or null input. `Origin.Inheritance` distinguishes it from a direct
declaration with concrete source, destination, and controlling directive paths.
A direct child replacement has its own effective origin; displaced ancestry
remains inspectable in history. An inherited relative resource keeps the
declaring source's passive `SourceContext`, not the child's directory.

The optional schema-version checker still runs once against the main input at
its original point. Inheritance cannot add, replace, or consume the protected
version path; an unrelated relationship inside an enclosing mapping does not
invalidate an unchanged version. Expansion uses the decoded projection
node/depth limits without charging copied values as acquired sources;
retained evidence uses the existing history-node limit. See the
[consumer example](consumers/README.md#named-sibling-inheritance) for one-load
usage.

### Source resource contexts

`Result.SourceContext(id)` and callback-scoped `Resolution.SourceContext(id)`
return a passive `ResourceContext` for a source with filesystem context, or
false for unknown and contextless IDs. The result query is nil-safe; the
resolution query follows the callback's lifetime guard. Queries perform no
filesystem access. Obtain IDs from exact origin locations rather than guessing
one context for a composite origin or an application-derived value.

`ResourceContext.FS` is rooted at the source's configured root, and `BaseDir`
is its slash-relative directory inside that root. This includes imported
sources, selected overlays, host files even without enabled imports, native
files, and byte, reader, or mapping inputs with an explicit context. CLI and
contextless synthetic inputs have no resource context. To open a relative
resource, join `BaseDir` with the application resource name using slash-path
operations and open that path through `FS`.

For a resource-valued field, query `result.Metadata().Lookup(fieldPath)`, inspect
its origin locations, choose a source ID only when the application's policy can
identify the intended contributor, then call `result.SourceContext(id)`. The
context belongs to that source even when the field survived from a lower-priority
underlay. A composite origin has no automatically selected resource base.

Native contexts borrow the caller's filesystem within the configured root.
Host contexts retain a canonical root path, not an open root handle; each
resource `Open` checks containment and opens a fresh handle that the returned
file's `Close` releases. Applications own and close the files they open.
Resources are application data, not configuration input, and do not consume
configuration byte or node budgets. The loaded configuration and metadata stay
usable if source files later change or disappear. Later resource operations
observe the backing filesystem then and may fail if files move, disappear, or
a caller-owned filesystem closes. This is not a snapshot or universal sandbox.

### Application schema versions

`VersionRule` declares a semantic path and non-nil checker. Before defaults or typed
decoding, the checker receives a `Version` preserving absent, explicit null, string,
or strict decimal-integer input. Strings and integers supply their exact spelling in
`Version.Text`; ConfigKit does not impose a machine-integer range. Fractional or
exponent numbers, booleans, mappings, and sequences produce `version.type`, and the
checker does not run. ConfigKit does not interpret the text or impose a universal
version path.

Integer versions use the strict grammar `^-?(0|[1-9][0-9]*)$`. Their spelling is
preserved exactly, including `-0`, and application policy decides which identifiers
are supported. A numeric version has no built-in `int64` range limit. Existing parser
policy still rejects leading zeroes, leading plus signs, hexadecimal notation, and
other non-JSON scalar forms before version checking.

At exactly the configured semantic path, accepted integer input can decode into a
plain Go string or named string without a text codec, including pointer wrappers.
This is a load-local exception: siblings and unrelated fields, dynamic values,
integer and float destinations, durations, `Secret`, and text codecs retain ordinary
strict decoding. The retained source and history stay numeric while typed projection
and effective YAML contain the resulting string; explanation records this through
the normal decoding reconciliation event. Fragment, overlay, and CLI ownership
protections apply before this conversion.

The checker reports safe, value-free errors and warnings through `Validation.Error`
and `Validation.Warning`; ConfigKit resolves primary and related paths to exact or
nearest input locations. Reported errors fail loading. If the checker returns an
error without reporting one, ConfigKit emits a generic `version.hook_failed`
diagnostic and suppresses the returned error text. Warnings alone are retained on a
successful result. Callback panics remain programming failures.

#### Shared source-version checks

`Options.SourceVersionMode` and `LayeredOptions.SourceVersionMode` select how
one `Definition.Version` policy applies across source documents. The zero value
`SourceVersionModeDefault` preserves the existing main-only check at its
effective path and existing protected-path behavior. `SourceVersionModeIfPresent`
checks each participating document that declares a version and permits absent
non-main declarations. `SourceVersionModeRequired` rejects a participating
document without a declaration. Enabled modes require a usable `VersionRule`
and checker; an unknown mode or unusable rule is an options error before source
access or callbacks.

In enabled modes, `Definition.Version.Path` is interpreted in each acquired
document's own coordinates, before placement and composition. The main document,
direct underlay documents, imported documents (including transitive and glob
matches), and the selected overlay participate. `NoSource`, absent optional
imports, empty underlay groups, defaults, resolution output, inheritance, and
synthetic CLI overrides are not documents to check. An absent declaration in an
actual main document still reaches the checker in if-present mode; required mode
rejects it first. With `NoSource`, the existing single absent-main checker
behavior remains in both enabled modes. If-present omissions in non-main
documents skip the checker. Explicit null is present and reaches the checker;
unsupported selected values produce `version.type`. No equality with the main
declaration or non-null rule is imposed. Acceptance establishes only that the
declaration is accepted by the consumer's checker; it does not establish that
each fragment conforms to a complete application schema. Final assembled
decoding and application validation still run once.

A header that is absent below a structural control follows the same omission
policy. Actual header/control overlap is rejected, and ordinary structural
validation and main-version ownership protection continue independently.

Accepted non-main declarations are consumed from their owned payload before
composition; their authored ancestors remain, even if empty. The main
declaration remains in the application. Ordinary release metadata should use a
different path. Enabled main protection uses the main input's effective path:
its `Destination` prefixed to the source-local `VersionRule.Path`. This protects
the remaining payload against imports, underlays, inheritance, and CLI
overrides. Thus a mounted main looks up a local header while ownership checks
protect its destination-prefixed application path. Main checker diagnostics and
declaration locations use source-local coordinates; ordinary payload
diagnostics continue to use effective paths. Default-mode diagnostics are
unchanged.

`SourceInfo.VersionCheck` is nil when enabled declaration processing did not
occur. Otherwise it records a source-local `Path`, the observed `Version` state
(`Present`, `Null`, and, when disclosed, exact `Text`), an actual or nearest
available `Location`, and whether text was redacted. A non-nil record with
`Present == false` records observed absence; it does not say the checker ran or
accepted a declaration. Absent evidence still identifies the full requested
header path, while its location identifies the nearest available source node.
During callbacks and in failed-load source records,
text is conservatively withheld because final classification may not be
available. Successful final source records reflect completed sensitivity
classification at the declaration's mounted path and ancestors, including
historical classification. Redaction empties `Text` while retaining presence
and null state. Returned copies are detached; evidence does not reopen inputs.
Consult [ADR 0013](adr/0013-shared-source-version-checks.md) for the accepted
decision and the [development guide](development.md#baseline-validation) for
the completed implementation and remediation records in Git history.

### Layered metadata

`Result.Metadata` describes every final canonical path. Queries retain explicit
key/index distinctions using actual effective and historical container shapes;
only parsed numeric tokens are resolved contextually. A descendant that never
existed is not historical merely because an atomic ancestor was replaced. Exact
surviving direct or inherited input paths are explicit and retain their source
layer and derivation; omitted paths have a default origin.
An explicit parent mapping does not make omitted children explicit. Null state and
available source coordinates are retained. `History` records the Go-zero or application
default baseline followed by exact base or overlay definitions, including
equal-valued input. Overlay replacement removes displaced descendants from effective
metadata while retaining their base definition and the parent replacement operation
in history. All metadata queries and returned nested records are detached.

`Metadata.HistoricalPaths` enumerates observed paths, including removed paths,
without reloading input or resolving unselected credentials. Consumers can use
it with `History` to validate ownership of declarations hidden by an overlay;
domain ownership rules remain application responsibilities. The method's Go
documentation in [metadata.go](../metadata.go) defines enumeration and path-ordering
details.

History is chronological: a path can have a synthetic default record, its exact
base definition, an overlay replacement, structural inheritance, and a recorded
resolution operation. A descendant
query uses the ancestor's path for a whole-container or whole-sequence replacement;
the descendant can therefore retain history after `Lookup` stops reporting it as
effective. Equal-valued input is still an explicit base definition, while a derived
value for absent input remains non-explicit.

Strict composition assigns contributed records their assembled sequence positions
in history. Across tiers, history follows the selected
[collection operation](#strict-document-composition), including append and
whole-record upsert. Incoming contributions use their assigned positions, and
replaced payload remains available as historical evidence.

A mapping assembled from the selected file and imports has one composite base origin
whose locations follow [scoped assembly order](#explicit-imports), not source-record
acquisition order. A mapping also visited by an overlay has an overlay-layer
composite origin with the lower contributing locations followed by the overlay
location; an empty mapping is still an explicit container even though it contributes
no children. Resolution origins are synthetic: they have layer
`resolution`, no fabricated file location, and a `derived_default` or `normalization`
operation. Reasons and dependency paths retain first occurrence order with duplicates
removed. Source records and IDs retain [acquisition order](#explicit-imports),
including the underlay ordering described in [layered loading](#layered-loading).

Inheritance adds no acquired source or precedence layer. `Origin.Inheritance`
lists concrete transfers from contributor paths through destinations, with
each step's controlling `extends` path and location. `History` retains source
applications at their original paths and projects parent contributions, child
replacements, and consumed controls at destination paths after source assembly
and before typed decoding. `HistoricalPaths` includes consumed directives and
displaced inherited descendants. A control can have history even when `Lookup`
reports it absent. Queries return detached records.

### Secret classification and effective rendering

Secret classification combines `configkit:"secret"` field tags with the optional
pure `Definition.Secret(Path) bool` predicate. Tags on concrete structs inside
supported dynamic values are retained in typed projections, including superseded
defaults and pre-resolution values. A tagged or predicate-classified
container makes all descendants sensitive. ConfigKit evaluates final and historical
paths, so replacing or deleting a definition does not erase its classification. The
predicate can add sensitivity but cannot declassify a tagged path.

Recorded resolution dependencies propagate sensitivity to their derived target.
Propagation reaches a fixed point, so record/map iteration order does not affect the
result and dependency cycles terminate. A dependency on a classified descendant or
on a container that contains one is treated conservatively as sensitive. Applications
must classify secret-producing transformations whose dependencies are not recorded.
Inherited value and control paths link source and destination classifications
through chains and nested copies; unrelated sibling fields are not classified
merely because another field is secret. Displaced historical evidence remains
redacted. Secret references are copied only as unresolved names; credential
lookup stays after decoding and application resolution.

`Result.WriteEffective` writes the frozen resolved snapshot as deterministic redacted
YAML. Mappings sort by decoded key string, sequences retain order, indentation is two
spaces with noncompact sequence indentation, line endings are LF, the document has no
start or end marker, and output ends with a newline. Nil pointers/maps/slices remain
`null`; non-nil empty maps and sequences remain `{}` and `[]`. A secret root is
replaced by the quoted string `'[REDACTED]'`. Redacted output is intended for
inspection and is not promised to be reloadable.

Rendering uses only a safe adapter tree built from the already completed canonical
typed projection. It does not invoke application YAML marshal hooks or text codecs;
all text-codec projection has succeeded before `Load` returns. Caller mutation of
`Result.Config` cannot change later output. Writer failures, including short writes,
are returned and may leave partial redacted output; internal dumper failures are
translated to a fixed safe error. There is no unredacted rendering option and
ConfigKit performs no implicit logging.

`Result.Config` is the explicit trusted boundary and contains ordinary typed values,
including secrets. Metadata records, diagnostics, and load errors contain no
configuration values. Explanation snapshots deliberately expose non-secret
configuration history after applying the redaction contract below. Applications
must still keep source labels, resolution reasons, and diagnostic messages value-free
because ConfigKit cannot infer and sanitize secrets embedded in arbitrary application
prose.

### Historical-value explanation

`Result.Explain(path)` returns a detached `Explanation` for any known effective or
historical path. Its effective value and each chronological before/after value are
`ValueSnapshot` records containing deterministic YAML text rather than mutable Go or
parser values. `Present=false` with empty YAML means absence or deletion; explicit
null is present and renders as `null`. The history starts with the canonical typed
application-default or Go-zero baseline, followed by relevant underlay and main-tier
contributions in precedence order, the selected overlay, and structural inheritance
events. Typed defaulting and decoding reconcile that source history before supplied
CLI records and audited net resolution operations. Within each strict tier, replay follows
[scoped assembly order](#explicit-imports), not source-record acquisition order.
Inheritance events include consumed or canceled `extends` input and equal-valued
child replacements. Overlapping parent/child resolution records
are rendered
as one net subtree event with combined reasons and dependencies, rather than
pretending that the recorded assignments ran in lexical order. Explicit equal
writes remain visible. When structural replay differs from the pre-CLI typed input,
a `ChangeDecoding` event records retained defaults and canonical decoding. CLI
contributions then retain their raw input representation while effective and
resolution values use the typed projection. This built-in inspection operation is
not accepted by `Resolution.Record`.

Base, overlay, and CLI history preserves the policy-validated input representation,
including a superseded value that could not have decoded into the final Go field.
Baseline, effective, and resolution values use canonical typed projection. A
decoding event bridges replayed input and canonical typed values, so intermediate
null/scalar replacements cannot incorrectly imply permanent default removal. Mapping
queries reconstruct each source or override contribution. A sequence snapshot
represents the collection at that point in replay, and sequence diffs remain
atomic. Descendant queries describe their presence and values before and after
containing sequence operations. Combination follows the
[collection operation tables](#strict-document-composition); upsert replaces
matching records in full. Removed paths remain explainable; unknown and malformed
paths return false.

Every snapshot applies the same historical tag, predicate, and recorded-dependency
sensitivity closure as effective rendering before YAML is exposed. A redacted scalar
is the quoted marker `'[REDACTED]'`, and a container snapshot has `Redacted=true` if
it contains a masked descendant. Typed `Secret` snapshots contain only the configured
environment name. `SecretReference` reports that name only for a successfully
resolved effective reference, and is omitted when its path is classified. Credential
values are never retained in explanation history.

`Result.WriteExplanation(output, path)` renders the safe structured result with the
canonical JSON-pointer path, canonical-versus-input representation labels, effective
source, composite locations, chronological values, resolution reasons and
dependencies, and successful secret-reference lookup attribution. Control characters
in paths, source identifiers and labels, reasons, dependencies, and reference names
are escaped. An unknown path returns a fixed safe error before writing anything;
writer errors and short writes are returned and may leave partial output. There is no
unredacted mode.

Retained history consists of immutable source applications and audited subtree
deltas, rather than a full tree copy per field. Shared nodes count once toward a
4,000,000-node retained-history limit; exceeding it fails loading with
`explain.limit_exceeded`. Replaying one requested path is performed on demand, and
the returned YAML strings and nested records cannot mutate later explanations.

Application commands should render the separate `SecretsReport` immediately after
an explicit `LoadSecretsDir` bootstrap, then use `WriteDiagnostics`,
`WriteEffective`, or `WriteExplanation` on the later load result as appropriate.
ConfigKit does not infer bootstrap outcomes from a secret reference and does not
bundle these independent reports into a command framework.

## Effective configuration diffs

`configkit.Diff(left, right)` compares two completed successful loads of the same
configuration type. `left` is the earlier state and `right` is the later state.
Applications load both sides with the sources and options they choose; to inspect
an overlay's impact, load the same base once without the overlay and once with it
selected. The [consumer guide](consumers/README.md#effective-configuration-diffs)
links to maintained examples. The public API returns safe changed-region records;
applications own their presentation.

Comparison uses the frozen canonical effective inspection values, not mutable
`Result.Config`, source text, provenance, or history. It does not reopen sources,
reread environment variables, or rerun defaults or callbacks. Equal effective
values produce no records even when sources, explicit-input status, defaults, or
resolution histories differ; for example, explicit zero and defaulted zero are
equal. Mappings compare recursively, sequences compare atomically, and scalar,
kind, or null/container replacements report their affected region. Records use
the existing comparator's canonical pointer ordering. `Present=false` with empty
YAML means absence; explicit null is present and renders as `null`.

Before either side is returned, sensitivity from both loads is applied to both
snapshots, including historical and dependency-derived classification. A
classification on either side protects the corresponding values on both sides.
`Redacted` reports masking, while `HiddenChanged` reports an actual concealed
value or presence change; it is not implied merely by redaction. A structural
replacement that prevents safe correspondence masks the complete reported
region, and an absent side remains absent. Record membership establishes that a
change occurred even when safe snapshots render identically. Literal
marker-looking input is ordinary input; consumers use the record fields rather
than interpreting marker text.

Typed secret references contribute their configured environment names to the
frozen comparison, never resolved credential bytes. Credential rotation behind
an unchanged reference is therefore not detected. The records contain changed
regions only, not paired full-document context; independently calling
`WriteEffective` on each result does not create cross-protected context. Nil or
uninitialized results return an error with no records. A failed `Load` retains
its normal nil-result/`LoadError` behavior, which callers handle before invoking
`Diff`.

## Typed secret references

Credential-bearing fields use the opaque `Secret` type and store an environment
variable name in YAML, such as `api_key: OPENROUTER_API_KEY`. `SecretEnv(name)`
constructs the same unresolved reference for application defaults or recorded
resolution. Names must match `[A-Za-z_][A-Za-z0-9_]*`; they are public
configuration and remain visible through `EnvName`, canonical projection, history,
and effective output unless the path is separately classified for redaction.

A non-pointer `Secret` is required. An absent, null, or invalid reference, or an
unset or empty environment value, fails with a safe `secret.*` diagnostic at the
configuration path. `*Secret` is optional only when nil; every non-nil reference
must resolve to a nonempty value. Secret references work in typed structs, maps,
pointers, slices, and arrays. A dynamic `any` value does not establish secret
semantics and cannot contain `Secret`; applications must use a typed slot.

ConfigKit resolves only the effective names after CLI assignment and recorded
application resolution, and before semantic validation. CLI overrides may supply
or replace a typed secret reference; lookup uses that effective name and retains
its effective source location for diagnostics. Each unique name is read with
`os.LookupEnv` once per load. Validation can inspect the credential through the
deliberately named `Value()` accessor and distinguish success with `IsResolved()`.
Later loads read the environment again, and cloned defaults are rebuilt from names
so a previously resolved result cannot retain a stale credential. `Load` never
writes to the environment.

Credential bytes are held only in the per-load lookup state and the trusted returned
configuration, and are returned only by `Value`. Secret formatting uses a safe
public-name representation, JSON marshals only the name, and projection, metadata,
histories, diagnostics, effective output,
and resolution snapshots never retain the credential. Validation remains read-only:
in addition to ordinary projected-value comparison, ConfigKit checks the private
resolved state against this load's lookup result. Validator mutation checks also
reject changes to projected field sensitivity, even when a replacement dynamic
concrete type preserves the same textual values. Applications must keep their
validation messages value-free and must not include `Value()` in diagnostics.

For example, a typed pipeline field is declared as:

```go
type PipelineConfig struct {
    APIKey configkit.Secret `yaml:"api_key"`
}
```

The executable `ExampleSecretEnv` demonstrates loading and inspecting the safe
reference state without printing its value.

## Explicit secrets-directory bootstrap

`LoadSecretsDir(directory)` is a separate, opt-in startup operation. It is never
called by `Load`, is not a YAML option, and does not participate in configuration
validation. Applications choose the native directory path and should invoke the
operation once before starting concurrent work. The selected path may itself be a
symlink; ConfigKit resolves it to an existing canonical directory before examining
its immediate entries.

The bootstrap enumerates at most 1,024 entries, including dot-prefixed entries.
Entries whose names begin with `.` (files, directories, or symlinks, such as `.git`
and `.gitignore`) are ignored before name or metadata validation, are never opened,
and are omitted from the report. Remaining entries are sorted by filename; each
must have an ASCII environment-variable name and be a regular, nonsymlink
file. Metadata for all remaining entries is validated before file content is read
or process state is changed. Existing environment variables win, including variables whose
value is empty; their files are not opened. Each eligible file is limited to 1 MiB,
and eligible content is limited to 8 MiB in total. ConfigKit removes exactly one
terminal LF or CRLF, preserves all other whitespace, accepts an empty value, and
rejects NUL bytes.

All eligible files are read successfully before application begins. Presence is
checked again immediately before each `os.Setenv`: a variable populated since
preflight is skipped. A write failure stops application without rollback, so prior
writes remain and later eligible entries are `not_applied`. This operation is not
atomic and does not provide compare-and-set semantics for concurrent environment
mutation.

`SecretsReport.Entries` returns detached, filename-ordered `SecretFileResult`
records with `loaded`, `skipped_existing`, `not_applied`, or `failed` status.
`ExistingEmpty` distinguishes an empty existing winner. `SecretsReport.WriteTo`
renders only quoted names, canonical paths, statuses, and that empty-state flag;
it never retains or renders file contents. Failures return the non-nil partial
report with a `*SecretsError` containing detached `secrets.*` diagnostics and no
raw filesystem or environment error chain. Applications should retain the report
alongside the later configuration result and explicitly present blocked and failed
files when appropriate; ConfigKit stores no implicit global report.

### Recorded application resolution

An optional `Definition.Resolve` callback runs after successful defaults, typed
decoding, and explicit CLI assignment. It receives the working configuration and a
per-load `Resolution` context. The context's
`IsExplicit` query uses surviving input presence, so explicit false, zero, empty, and
null are never confused with absence or static defaults.

`Resolution.InputInfo(path)` returns detached effective metadata from the fixed
post-decode entry state, including explicit CLI overrides, before application
resolution runs. It includes retained static defaults, preserves explicit null, and
uses the same exact semantic path matching as `Metadata.Lookup`; missing,
malformed, and key/index-shape-mismatched paths return false without diagnostics.
Its `Explicit` value agrees with `IsExplicit` for every found entry path. Resolution
changes, additions, removals, and moves do not alter this entry view.

`Resolution.Source(id)` returns a detached record for an already acquired source in
the same load: the main source, imports, selected overlay, and supplied CLI records
where applicable. Empty and unknown IDs return false. It never reloads input, reads
the filesystem, resolves symlinks, or looks up credentials. A source path exists
only when the acquired source has one; defaults and byte, reader, and CLI sources do
not acquire an invented file path. Applications select a base directory from exact
origin locations and these source records according to their own policy.

Both queries are callback-scoped like the rest of `Resolution`: nil, zero, or
retained contexts panic with the fixed closed-context failure outside the callback.
Returned field and source records are detached and remain usable after callback
exit; changing them cannot affect later queries, results, or another load.

`Default` assigns and records a derived value only when its path is not explicit.
`Normalize` always assigns and records. For transformations performed directly by
application code, `Resolution.Record` declares a derived-default or normalization
operation; a recorded container covers changes to its descendants. Records require
a schema-valid path, valid kind, safe nonempty reason, and schema-valid dependency
paths. Repeated records at one path are coalesced: reasons and dependencies retain
their first occurrence, the last helper value is authoritative, and any normalization
makes the combined operation a normalization.

ConfigKit projects the configuration before and after the callback. Every net change
must have an exact or ancestor record. Helper-assigned values must agree with the
final value at their declared path, which supports ordinary map-entry copy/writeback
while detecting forgotten writeback or unrelated targets. Invalid paths, kinds,
values, records, helper targets, and unrecorded mutations fail the load. Equal-valued
derived assignments may still record application intent even when there is no
structural change.

`Resolution.Error` and `Resolution.Warning` attach exact or nearest available input
locations. Reported errors fail the load; warnings survive on a successful result. A
returned resolver error adds one generic `resolve.hook_failed` error only when the
callback did not report an error, and its text is never exposed. Callback panics are
programming failures. Resolution and validation contexts are callback-scoped; using
a nil, zero, or retained context outside its active callback panics with fixed,
value-free text.

Successful records update effective origins and append detached metadata history
without changing input explicitness. Removed paths leave effective metadata but
retain history. ConfigKit privately snapshots each recorded before/after subtree for
later safe explanation; caller mutations of `Result.Config` cannot alter those
snapshots.

### Semantic validation

An optional `Definition.Validate` callback runs after successful resolution and
receives the resolved configuration plus a per-load `Validation` context. It can
collect multiple independent `validate.error` and `validate.warning` reports.
Warnings continue the lifecycle and are returned on success; any error returns a nil
result with the full safe diagnostic and source report.

Primary and related paths must be valid for the application schema. Omitted declared
fields and missing string-map or dynamic keys are valid report targets; unknown
struct fields and malformed paths produce `validate.invalid_path`. ConfigKit attaches
the exact surviving input location when one exists, otherwise the nearest input
container location, or no primary location for defaults-only synthetic state.

Validation is read-only. ConfigKit projects the configuration immediately before and
after the callback and returns `validate.mutation` at every detected net-change path,
including nested mutable state. A value that can no longer be projected is also a
mutation failure. No mutated or otherwise invalid configuration is published.
Callback panics remain programming failures, and retained contexts use the same
fixed closed-context panic as version and resolution callbacks.

Only schema-bound exported state participates in loading, resolution audit,
validation audit, metadata, and snapshots. Ignored or unexported ordinary fields are
zero in returned detached values. Applications retain responsibility for domain
rules and runtime initialization; validators should not start clients or depend on
live services as a substitute for configuration validation.

## Strict document composition

Ordinary `Load` applies strict document composition to its parsed source scope;
`Options.Imports` adds imported documents as peers in that scope. Application
defaults do not participate. With imports disabled, the single source still has
its controls extracted and any declared or registered collection identity is
validated, but there are no imported peers to combine.

| Existing node | Incoming node | Result |
| --- | --- | --- |
| Absent | Absent | Absent |
| Absent | Present | Retain the incoming node |
| Present | Absent | Retain the existing node |
| Mapping | Mapping | Recurse by mapping key |
| Any other pair of present nodes | Any other pair of present nodes | `compose.conflict` |

Every non-mapping overlap is a conflict, even when values are equal, except for
sequence/sequence contributions at a destination with effective identity from an
application registration, a current YAML `key` declaration, or retained lower
policy. Nulls remain atomic conflicts.
Sequences without effective identity conflict when repeated and are never joined
by index. Disjoint entries under shared mappings compose, making stable keyed
mappings the natural representation for independently contributed entries.

`Definition.SequenceIdentities` contains registrations created by
`SequenceIdentityAt` or `SequenceIdentityMatching`. The exact form accepts a
mapping-only `Path`; the matching form accepts literal-key and any-mapping-key
segments created by `SequenceLiteralKey` and `SequenceAnyMappingKey`. A literal
`"*"` key is distinct from the wildcard constructor. Parsed numeric path tokens
remain literal mapping keys; explicit index segments and reserved `configkit` or
mapping-valued `extends` segments are invalid. The root selector is harmless under
the mapping-root schema. Selector segments consume one mapping level and match a
complete destination; an any-mapping-key segment matches exactly one key, not an
arbitrary prefix or suffix. Each registration requires one or more literal,
direct field names; compound field sets are normalized as sets, and field names
are not otherwise normalized. The same field-set and scalar-equality rules apply
to YAML `key` declarations.

At each selected path, records must provide every effective identity field as a
scalar string, number, boolean, or explicit null. Field names form a set, so
order and repetition do not affect compound identity. The type-tagged tuple
preserves category distinctions: strings compare exactly, including empty
strings; numbers, booleans, and null remain distinct categories. Finite numbers
compare by exact mathematical value, so `1`, `1.0`, and `1e0` are equal, as are
`0` and `-0`, without floating-point rounding. Distinct tuples append in source
and list order; duplicate tuples conflict even when records are otherwise equal
or disjoint. Malformed records produce `compose.invalid`; duplicate identities
produce `compose.conflict`, with paths and source locations identifying a
duplicate without exposing identity values. Missing paths and non-sequence values
do not fail by themselves. Equivalent overlapping registrations coalesce.
Incompatible overlaps and invalid registrations produce `options.invalid` before
source acquisition. Unused and root registrations are allowed and have no effect.

Registrations provide identity for strict composition, sequence operations,
inheritance lookup, and final validation. Dynamic selectors apply at concrete
strict-composition, inheritance, and final-result destinations without requiring
the application to enumerate mapping keys. Each strict tier has independent
duplicate-identity state among its peers; layered precedence and inheritance
then apply the collection operation described below.

Final registration checks run in both ordinary and layered loads after effective
secret lookup and before application semantic validation. They examine actual
final typed sequences, including values introduced by defaults, CLI, or recorded
resolution; they do not require path existence or sequence shape. Diagnostics
use concrete collection and item paths and available source locations, without
exposing identity or credential values. The guarantee ends when loading publishes
successfully and does not cover later caller mutations.
YAML-only identity is enforced during structural composition, precedence, and
identified inheritance; it carries no final-result guarantee after typed
defaults, CLI, or application resolution. Applications that need that guarantee
register the identity constraint.
The [maintained catalog examples](consumers/README.md#identified-prompt-sequences)
show strict peers, precedence operations, inheritance, dynamic destinations, and
the distinction between YAML-only declarations and application registrations.
When an enabled source-version mode is selected, declarations are checked in each
source's own coordinates before placement and composition. Accepted non-main
headers are consumed before strict composition and cannot supply a registered
record identity. Final typed decoding and application validation still run on the
completed configuration.

Ordinary sources, layered inputs, and selected overlays may declare collection
behavior under `configkit.sequences` at the document root. Each entry has a
required literal `destination` and may specify `key` as one string or a nonempty
list of strings, plus a `mode` of `replace`, `append`, or `upsert`. These
controls are removed before application decoding. A declaration destination is
relative to that source's placement, so a mounted fragment's declaration is
mounted with its values. Declarations from strict peers must agree on identity
fields and on any explicitly supplied mode; omission expresses no mode
preference. They apply even when the declaring source has no value at that
destination: a key validates an existing retained sequence or remains pending
until records arrive, while a mode-only declaration affects later combinations
without recombining earlier tiers. A key supplies the identity used by `append`
and `upsert` and is checked against any application registration.

Strict peers at an identified destination first combine distinct records in
source and list order. Duplicate identities among peers or within one
contribution fail under every mode. After strict composition, precedence and
inheritance use one operation for the assembled incoming collection:

| Mode | Behavior at source-precedence and inheritance boundaries |
| --- | --- |
| `replace` | Replace the complete lower sequence with the assembled higher sequence. |
| `append` | Retain lower records and append higher records; duplicate identities fail. |
| `upsert` | Replace matching records in full at their lower positions, retain unmatched lower records, and append new identities. |

Upsert replaces the complete matched record; it does not recursively merge
record fields. Append and upsert require identity. With no explicit mode in the
current tier, an explicit lower mode persists; if none exists, effective identity
selects upsert and its absence selects replace. Inferred modes do not persist.
Explicit modes persist at their destination until changed by a higher explicit
mode, independently of YAML identity lifetime. A declaration alone does not
clear or recombine a collection.

The presence and value kind at a precedence or inheritance boundary determine
whether a collection changes:

| Higher contribution | `replace` | `append` / `upsert` |
| --- | --- | --- |
| Absent | Leave the lower value unchanged. | Leave the lower value unchanged. |
| Explicit `[]` | Clear the sequence. | Contribute no records. |
| Explicit null or another non-sequence value | Replace the lower value, subject to application schema. | Error; the operation requires a sequence. |

An absent lower value can be initialized by an incoming sequence under any mode.
A present lower null or other non-sequence value is not absence: append and
upsert cannot combine with it. An empty sequence still requires a valid mode and
identity where applicable. Null used as an identity component inside a record is
distinct from null used as the collection value.

Declarations govern ordinary strict source composition, including loads with
imports disabled, and each strict scope in layered loading. Before composing a
higher tier's peers, ConfigKit makes applicable lower identity and explicit mode
available; after peer composition, the tier applies its resolved operation
against the lower collection. Selected overlays use the same collection policy
after the base source graph is composed. During mapping-sibling inheritance, a
parent's YAML identity can follow its collection to the child destination, while
the parent's explicit operation stays at the parent destination. The child
destination selects its operation; absent one, effective identity implies
upsert and its absence implies replace. Whole-sequence replacement discards
lower YAML identity and uses the replacing scope's key declaration, if any;
without one, the collection is unkeyed unless an application registration still
supplies identity. Replacement discards policies owned by removed records, while
an independently declared explicit mode and application registration remain
applicable. Item-local
policies follow their owning records: retaining a record retains its rules, while
whole-record upsert replacement discards the displaced record's rules. Matching
identity or reusing an index does not transfer them. These policies do not change
CLI behavior: CLI sequence assignments replace whole sequences after decoding.
A root `configkit.imports` declaration still requires `Options.Imports`;
`NoSource` has no declaration scope.

An empty mapping contributes no child definitions. Two empty mappings compose, and
an empty mapping neither clears nor conflicts with another mapping's children.
Shared mapping containers retain all contributing value locations as their composite
base origin. A conflict reports the first claim in
[scoped assembly order](#explicit-imports) as its primary location and every
remaining claim as related locations in the same order. Assembly order never
authorizes replacement. Composition clones the resulting tree and its ordered
claim records rather than mutating or aliasing an input document.

## Selected overlays

`Options.Overlay` selects at most one overlay by name. Names must match
`[a-zA-Z0-9_-]+`; they are not paths. The empty name and exact lowercase name
`default` select no overlay, and `overlays/default.yml` is never loaded implicitly.
Every other name resolves only to `<configuration-root>/overlays/<name>.yml`, with
no extension fallback, case normalization, inheritance between overlay files,
or overlay stack. A selected missing overlay is an error.

With `NoSource`, the main input stays absent and the explicit filesystem context
supplies the overlay root. The overlay is recorded as the first acquired source
when no other source exists. Host contexts may use the root-only form described
under [single-source loading](#single-source-loading); native contexts retain
their complete namespace, base, and root requirements.

The overlay is a separately policy-validated YAML mapping. It is loaded after the
base import graph is composed and its schema version is checked, but before defaults
and typed decoding. Overlay bytes, nodes, and source records count toward the same
aggregate limits as base inputs. Its canonical path must remain within the configured
root and must not duplicate a file already present in the base graph.

Collection destinations follow the
[collection operation and presence rules](#strict-document-composition) above.
The following table describes ordinary overlay behavior elsewhere:

| Existing lower node | Incoming overlay node | Result |
| --- | --- | --- |
| Absent | Value | Retain the overlay value |
| Value | Absent | Retain the lower value |
| Mapping | Mapping | Recurse by mapping key |
| Any other pair of present nodes | Any other pair of present nodes | Replace the entire lower node |

Under these ordinary rules, scalars, sequences, and null are atomic. An overlay
sequence replaces rather than extends a base sequence, `[]` clears it, null is an
explicit replacement, and cross-kind replacement is allowed when the final
application schema accepts it.
Replacing a subtree removes all displaced descendants from effective presence and
origin data; history retains the displaced definitions and identifies the parent
replacement. Surviving overlay paths have overlay origins, while recursively shared
mapping containers retain their contributing locations.

An overlay may contain only root `configkit.sequences` declarations; overlay
imports and nested directives are rejected. Sequence declarations are removed
before application decoding. At declared or registered sequence destinations,
`replace`, `append`, and `upsert` use the same identity and policy rules as source
composition. Other overlay sequences remain atomic replacements. An overlay may
not contribute the configured schema version or replace an ancestor of that
version with a scalar or null. These rules apply even if the supplied version
text would be unchanged. Overlays cannot mask a base parse, composition, or
version failure because those checks run first.

## Explicit CLI overrides

`Options.Overrides` accepts only records the application knows were explicitly set
by its CLI parser. Each `Override` supplies a non-root mapping `Path`, a detached
`Value`, and a required safe display `Flag` such as `--workers`; ConfigKit has no CLI
parser dependency. Parser defaults must not be submitted as overrides, and `Flag`
must never contain the raw argument or configured value. General environment
variable overlays are an explicit non-goal: environment access is reserved for
typed secret references.

Overrides are prepared after source assembly and inheritance, then assigned after
typed defaults and decoding and before application resolution. The order is:
defaults and decoded sources < CLI < application resolution. A CLI mapping
recursively overlays the decoded mapping; a CLI sequence replaces the complete
decoded sequence regardless of source collection operations. Inheritance and
typed defaulting do not run again after CLI assignment. Earlier parsing,
composition, inheritance, and decoding failures remain errors even if an override
would replace the failing value. Recorded resolution, typed secret lookup, and
semantic validation use the effective post-override values. Overrides also work
with `NoSource`.

Values are converted directly into ConfigKit's owned tree without YAML parsing or
string coercion. Supported values are `nil`, built-in strings and booleans, all
built-in signed and unsigned integer widths except `uintptr`, finite `float32` and
`float64`, and recursive `[]any`/`map[string]any`. Strings stay strings; duration and
secret fields therefore receive string values. Nil is explicit null; empty strings,
false, and zero remain intentional; empty sequences replace whole lists; and mapping
values recursively overlay existing mappings. Map keys are sorted canonically and
sequence order is preserved.

All records are detached and validated before any is applied. They cannot create
`configkit` or `extends` controls. Root and explicit
sequence-index paths are invalid; parsed numeric tokens remain mapping keys. A path
may create missing mapping ancestors but cannot traverse an existing scalar, null,
or sequence. Duplicate and ancestor/descendant-overlapping records are rejected.
Neither paths nor values may create `configkit`, and no record may overlap the
configured schema-version path. Unknown fields and effective type mismatches remain
ordinary strict-decoder errors.

At most 1,024 records are accepted, and each record creates one synthetic CLI
source after file/overlay sources in supplied order. Override sources share the
1,024-source ceiling. A record has an 8 MiB string-byte limit; override strings and
source bytes share the 64 MiB aggregate limit. Synthetic mapping keys and values
count toward the existing 1,000,000-node and depth-256 ceilings. Cycles, named or
unsupported types, nonfinite floats, and limit excesses fail with safe `override.*`
diagnostics. Diagnostics and provenance identify the safe `Flag` source, never raw
arguments or values.

Override origins are explicit CLI origins. Resolution receives CLI values with
their explicit-presence and source information. Absence-based defaults preserve
those values, while recorded normalization and derivation may change them. History
retains replaced source definitions and records equal-value replacements as well
as structural changes. Secret lookup and validation use post-resolution values
with their effective source locations.
Caller-owned maps and slices are snapshotted during loading and must not be mutated
concurrently with the call. The executable `ExampleOverride` shows translating a
parser's explicit-set state into records while preserving an explicitly supplied
zero.

## Explicit imports

Imports are opt-in. With `Options.Imports` enabled, a document may use its root
`configkit` mapping for `imports` and `sequences` controls. Ordinary `Load` may use
`sequences` without enabling imports, but an `imports` field still requires
`Options.Imports`. ConfigKit removes recognized root controls before application
decoding. Each imports sequence item is a mapping with exactly one `path`,
`dir`, or `glob` selector. `path` and `dir` accept a string or a list of strings;
`glob` accepts one path pattern string. The optional `required` boolean controls
whether an empty selection or a missing named path is an error. Its default is
false, so missing imports are absent unless `required: true` is specified. The
removed `optional` field is invalid. `destination` is an optional sequence of
strings. Every destination string is one literal mapping key and follows the
input-destination validation rules above; quote numeric-looking keys so YAML
reads them as strings. Null, scalar destinations, and non-string members are
`import.invalid`, as are all other unsupported shapes and fields. A `configkit`
key below the root is invalid. With imports disabled, root imports and nested
`configkit` keys remain `import.disabled`.

The directive key is `configkit`, not the `_configkit` spelling used in v0.1.0.
There is no legacy alias: `_configkit` is ordinary application data, subject to
normal schema validation, and does not trigger imports. When upgrading, rename
directive keys in root files and imported fragments, and rename any application
fields previously named `configkit` to avoid the newly reserved key.

Path declarations use slash-separated relative paths. Absolute paths and network
URLs are invalid. Each path resolves from its declaring file's canonical directory.
An empty `path: []` selects nothing; it succeeds even when `required: true` because
there are no named targets to be missing. A nonempty required path list verifies
every named target, including targets later excluded from selection.

Directory declarations search the listed literal directory roots. Missing roots
contribute no candidates; supplied inaccessible or invalid roots fail. `pattern`
accepts one basename pattern or a list, matching filenames only. Omission selects
all filenames; an explicit empty list selects none. `exclude` accepts one basename
pattern or a list and removes matching final candidates. Patterns use Go's
case-sensitive `path.Match` grammar. `recursive` enables descendant traversal;
`depth` is valid only when recursion is effective, with direct files at depth 1 and
depth 0 or omission meaning unlimited traversal subject to library budgets.
`follow` controls discovered symlink entries and `hidden` controls discovered
names beginning with a dot. Both default to true. Explicit roots retain their
normal access behavior.

Glob declarations use the slash-based Go `path.Match` grammar: `*`, `?`, character
classes, and backslash escapes are supported, while recursive `**` syntax is
invalid. Backslashes escape pattern characters rather than acting as directory
separators. ConfigKit validates the complete pattern before filesystem traversal,
walks only directories selected by preceding segments, and reads wildcard
directories in bounded batches. Every examined entry consumes the separate visit
budget, including nonmatches. Literal components are explicit path components;
`follow` and `hidden` filter only wildcard-discovered entries. `exclude` filters
final candidate basenames.

`Options.ImportDefaults` and `LayeredOptions.ImportDefaults` provide one
application-owned load-wide policy for enabled imports in the main source,
underlays, and transitive imports; they do not apply to selected overlays or
discover sources by themselves. Each applicable field resolves independently:
an explicit declaration value wins, then the application default, then the
library default (`required: false`, `follow: true`, `hidden: true`,
`recursive: false`, and unrestricted filenames). A declaration's `pattern`
replaces rather than extends the default pattern. Nil `Pattern` leaves directory
filenames unrestricted; a non-nil empty slice selects none. Defaults do not make
a field applicable to selectors that do not support it, and load calls snapshot
the supplied pointers and slice before source acquisition.

A source's `configkit.sequences` declarations apply independently of import
enablement; a `configkit.imports` declaration still requires `Options.Imports`.
Sequence identities can be supplied either by these YAML declarations or by
application registrations. The [strict composition contract](#strict-document-composition)
defines their agreement, operation, and lifetime rules.

Acquisition visits declarations in authored order and each declaration's selected
paths in namespace-path lexical order, preserving source-record and version-callback
order. Ordinary mapping contributions in a main tier are assembled dependency-first:
ConfigKit selects the lexically smallest stable source key among the contributions
whose imports are complete. Independent branches may interleave, and the main body
may be assembled after its imports. Sequence-item contributions in that tier are
nodes in the same schedule; each selected item first assembles its own strict
mapping scope, with helpers before the bodies that import them. Those local scopes
do not use application collection policies and are not flattened into the outer
schedule. Item positions and diagnostic paths follow the resulting assembly order.
Each underlay group likewise acquires its inputs before dependency-ready assembly
in one strict scope. Independent groups remain separate precedence tiers; at a
declared or registered destination, the assembled higher sequence uses the
resolved collection operation against the lower sequence. Other sequences retain
whole-value replacement between tiers. Assembly order never permits replacement.

Acquisition order and resolution order serve different purposes: acquisition
assigns source records and callback order, while the shared scheduler determines
composition order among currently dependency-ready contributions. Inheritance
uses the same eligibility rule within each destination-local sibling scope; it
does not reorder the authored output mapping. The private namespace and stable-key
construction is described in the [import implementation overview](internal/imports.md).

An empty selection is allowed by default. `required: true` makes an empty directory
or glob selection fail with `import.missing`; it also makes missing named paths
fail. Requiredness does not suppress invalid paths, containment failures,
directories selected as files, permissions, parse errors, or composition errors.

For inspection, `SourceInfo.ImportSelections` records one outcome for each
completed optional declaration whose entire selected set is empty. The record is
attached to its declaring source, keeps the declaration path local to that source,
identifies the declaration mapping (including its `SourceID`) and selector, and
uses outcome `none`. Records remain in declaration order and are available
through the existing result, load
error, and callback source snapshots. They do not enumerate candidates, explain
why a path was absent, or describe partial omissions. Empty `path: []` succeeds
regardless of requiredness: it produces a `none` record when effectively optional,
and no record when effectively required. Snapshots detach the record paths and
slices.

Each declaration may include `placement: mapping` or
`placement: sequence-item`. Omission means `mapping` for that declaration,
including every match of a glob; placement is not inherited. Sequence-item
placement requires a nonempty effective destination and wraps the imported
mapping as one element of a sequence there. It does not split source sequences
or infer identity fields. Repeated sequences combine only when the destination
has an applicable YAML identity declaration or application registration; source
placement remains independent of collection identity and operation. Sequence
records retain source locations through placement and history; indexes describe
their assembled positions while source coordinates continue to identify the
declaring files.
Header/version policy retains its existing owner and timing. Application
registrations also enforce identity on the final effective typed configuration;
YAML-only declarations are enforced at structural combination and inheritance
boundaries. The [configuration contract](#strict-document-composition) defines
these validation boundaries.

An item import first opens a local mapping scope. Its body and ordinary helper
imports compose within that record. An omitted helper destination or
`destination: []` places the helper at the record root; other destinations are
relative to the immediate importing source. A helper may supply fields such as
the record's `id`. An explicit nested `sequence-item` import creates a sequence
inside the record, and its placement is independent of the enclosing item.
At the record root, an item import must name a nonempty nested destination.
Physical paths continue to resolve from the file that declares each import.
An actual empty mapping contributes an empty item, while an unmatched declaration
with default or explicit `required: false` contributes nothing.

For host files, the configuration root defaults to the selected file's canonical
directory. A host file may supply `FilesystemContext.Root` to widen that root but
may not supply `BaseDir` or a native `FS`. Byte, reader, and mapping inputs with host context
must provide both `BaseDir` and `Root` as absolute existing directories.
For `Filesystem.File`, the default root is `.` in its namespace and the base is
the selected file's directory. An explicit context must use the same `Filesystem`
namespace and may set `Root` to contain the file and its base; it may not supply
`BaseDir`. Byte, reader, and mapping inputs with native context provide both `BaseDir` and
`Root` as slash-relative `fs.ValidPath` directories in `FS` (`.` means its root),
with the base inside the root. Context paths are checked even when imports and
overlays are inactive. Diagnostic labels, including URI-looking labels, never
imply filesystem paths. Nonregular targets are rejected before opening to avoid blocking on FIFOs;
post-open regular-file checks remain in place under the stable-input assumption.
ConfigKit canonicalizes symlinks,
checks containment with relative paths, and opens the canonical root-relative target
through a library-owned root handle. Contained relative or absolute symlink targets
are accepted; escapes are rejected. Canonical paths define identity, so hard links at
distinct paths remain distinct sources. Native filesystems use slash-path
operations. Where the supplied `fs.FS` implements `fs.ReadLinkFS`, ConfigKit uses
that capability to inspect and resolve discovered links; otherwise it follows
the filesystem's ordinary `fs.FS` behavior. ConfigKit does not require every
native filesystem to expose link inspection, and native filesystem behavior is
not a general sandbox guarantee. Imports stay in their selected namespace and
root.

An active-stack repeat is `import.cycle` and reports the complete cycle's import
sites. A previously visited canonical path is `import.repeated` and reports both
inclusion routes, including diamonds and aliases. Each accepted file is read once.
Imported fragments may contain their own root directives, but may not claim the
application's schema-version path or replace one of its ancestors; the version must
come from the selected source. No directory is scanned implicitly, including one
named `conf.d`, and ConfigKit performs no remote fetching.

## YAML documents

The private YAML boundary currently accepts exactly one nonempty document whose root
is a mapping. Comments and block or flow collections are supported. It rejects:

- empty input or an empty document;
- every additional document, including an empty one;
- duplicate decoded mapping keys;
- non-string mapping keys;
- explicit tags, anchors, aliases, and merge keys;
- a quoted or plain mapping key equal to `<<`; and
- input exceeding the fixed depth or node limits.

Mapping key comparison is exact: it performs no case folding or Unicode
normalization. The private document tree retains declaration order, exact validated
number spelling, source identifiers, and separate key/value line and column
locations. It contains no upstream parser nodes or types.

Root depth is zero. Input depth is limited to 256 and each parsed input tree to
1,000,000 nodes, counting mapping keys as input nodes. The upstream parser also has
a private calibrated depth guard; passing that guard never bypasses ConfigKit's tree
limits.

## Scalars

Scalar meaning is determined by ConfigKit rather than upstream coercion:

| Input form | Meaning |
| --- | --- |
| Quoted, literal, or folded scalar | String |
| Plain `true` or `false` | Boolean |
| Plain `null`, `~`, or an empty mapping value | Null |
| JSON decimal number | Number, retaining its exact spelling |
| Date/time-looking or other ordinary plain text | String |

JSON numbers permit an optional minus, a zero or nonzero decimal integer without a
leading zero, an optional fraction with at least one digit, and an optional decimal
exponent. Hexadecimal, octal, digit separators, leading plus, leading/trailing-dot
forms, incomplete exponents, and non-finite values are rejected. Legacy boolean
words (`yes`, `no`, `on`, and `off`) and non-lowercase variants of the reserved
boolean/null words are rejected unless quoted.

Mapping keys use the same scalar classification and must classify as strings. Quote
keys that otherwise classify as a boolean, number, or null.

## Application schema binding

ConfigKit builds a private reflection descriptor graph for each schema operation;
it keeps no global mutable descriptor cache. A configuration root must be a struct
or string-keyed map. Nested supported shapes are:

- booleans, strings, signed and unsigned integers, and finite floating-point types;
- `time.Duration`, typed `Secret` references, and paired
  `encoding.TextMarshaler`/`TextUnmarshaler` scalars;
- pointers, structs, string-keyed maps, slices, fixed arrays, and `any`; and
- recursive type definitions whose eventual application values are acyclic.

Named scalar types retain their Go identity. `uintptr`, complex numbers, functions,
channels, unsafe pointers, non-string map keys, and nonempty interfaces are not
configuration types. A custom text scalar must provide both directions on its value
or pointer method set. Building a descriptor only inspects method sets; it never
executes application methods or parser-specific hooks.

Exported struct fields use an explicit `yaml` name or the lowercase Go field name.
Unexported fields and `yaml:"-"` are ignored. `omitempty` is accepted but does not
change binding or projection. An explicit `yaml:",inline"` flattens a struct
or pointer-to-struct, or designates at most one string-keyed inline map for otherwise
unmatched names. Ambiguous names, named inline fields, multiple inline maps, and
unsupported or malformed tag options are schema errors. Inline pointers must
terminate in a struct; pointer-to-map and recursive pointer-only inline declarations
are rejected before the defaults factory runs. Embedded fields are ordinary
named fields unless explicitly inlined. `configkit:"secret"` is the only current
ConfigKit field tag and marks its bound field or container for secret classification
and redacted inspection.

Schema path traversal uses these same bindings. Explicit key segments address
mappings, explicit index segments address sequences, and parsed numeric pointer
tokens become indices only in a sequence. Sequence indices must use canonical
nonnegative decimal spelling: no signs or leading zeroes except `0` itself.
Tokens such as `+0` and `-0` remain literal keys in mappings but do not identify
sequence elements. Fixed-array indices are bounds checked. Dynamic `any` and inline-map entries
remain schema-valid for application-defined descendants, but dynamic values cannot
introduce typed secret references.

Descriptors are reused by typed decoding, projection, and cloning so those operations
cannot acquire different field bindings.

### Canonical module options

Applications with a known module catalog represent each configured item as an object
with an explicit module identifier and an options mapping. Ordinary typed structs,
including a small union of optional option fields, use the same strict binding above;
application validation decides which fields and combinations apply to each module.
String shorthand, a library module registry, and custom structural decoder hooks are
not supported alternatives.

When a runtime-selected catalog genuinely requires dynamic options,
`map[string]any` retains ConfigKit's canonical dynamic scalars. In particular,
integers are `int64`; application validation owns allowed keys, value constraints,
range checks, and conversion to runtime `int` or catalog objects. ConfigKit does not
provide generic subtree conversion, and consumers must not add a YAML marshal/decode
round trip to exercise supported behavior. The
[consumer adoption guide](consumers/README.md) contains compiled typed and dynamic
examples.

## Defaults and typed decoding

The private decoder starts from an independently owned clone of the application's
typed defaults and applies only explicitly present input nodes. Defaults therefore
never become explicit, while equal-valued input remains explicit. Successful decode
produces detached typed state, baseline and final canonical trees, an input-presence
index, and default/input origins for the later public loading lifecycle.

| Configured input | Effect on the typed baseline |
| --- | --- |
| Absent field | Retain the default |
| Explicit false, zero, or empty string | Replace the default |
| Partial mapping or string-keyed map | Recurse and retain omitted default children or entries |
| Empty mapping | Retain existing mapping children |
| Sequence, including an empty sequence | Replace the entire sequence |
| Pointer object | Recurse into a non-nil default; otherwise allocate an object whose omitted children have ordinary Go zero values |
| Allowed null | Replace the value with nil and suppress its default |

Null is allowed only for pointers, maps, slices, and dynamic `any` values. It is an
error for scalars, structs, and arrays. New map entries and sequence elements begin
from their Go zero values; ConfigKit does not invent entry-level defaults.

Decoding is strict at every typed depth. Unknown struct fields, type mismatches,
invalid nulls, numeric overflow, and arrays with the wrong length are errors at the
original semantic path and source coordinates. String-keyed maps accept arbitrary
keys but preserve the strict schema of their values; inline maps accept unmatched
keys according to their element descriptor. Multiple independent field failures are
collected without publishing the partially decoded object.

Booleans and strings are not generally coerced. At exactly a configured
`VersionRule.Path`, strict decimal-integer input may decode into a plain `string`, a
named string without a text codec, or pointer wrappers around those types. The
exception is load-local and does not apply to unrelated fields or loads without a
version rule. Integer destinations require integer syntax
and a value in the destination type's range. Float destinations accept finite JSON
numbers representable by their destination precision. Dynamic numbers become
`int64` for integer syntax and `float64` for fractional or exponent syntax, with
range and finiteness checks. Named scalar types retain their identity. Durations
decode only from strings accepted by `time.ParseDuration`; integers are rejected.
Paired text scalars decode only from strings through `encoding.TextUnmarshaler`, and
codec failures are classified without exposing application error text or input
values.

This decoder is available only through `Load`; ConfigKit exposes no generic document
conversion API.

## Typed projection and ownership

ConfigKit can internally project a schema-valid application value directly into its
owned canonical tree without YAML serialization. Projection includes every bound
field, including ordinary zero values; `omitempty` has no effect. Struct fields keep
binding order, dynamic map keys sort by decoded string, and sequences keep their
order. Integers retain exact base-10 values, finite floats use their destination
precision, durations use `time.Duration.String`, and paired text scalars use their
canonical marshaled text.

Nil pointers, maps, slices, and dynamic values project as null. Non-nil empty maps
and slices remain empty containers, preserving the distinction. Recursive acyclic
objects are supported. Cyclic values, non-finite floats, unsupported dynamic concrete
types, invalid codec output, and depth/node-limit violations fail with safe schema
classifications that omit codec error text.

The same descriptor graph creates detached typed copies. Pointers, maps, slices,
arrays, nested dynamic values, and shared references are recursively copied; shared
pointer identity is not preserved as configuration meaning. Paired text scalars are
rebuilt by unmarshaling their canonical text into fresh values, detaching private
backing storage. Ignored and unexported struct fields are zeroed rather than copied.

Private structural comparison operates on canonical trees. Mapping additions,
removals, and descendant replacements retain semantic paths. A sequence is atomic:
any sequence change reports the sequence path and never invents correspondence
between old and new indices. Cross-kind and null/container changes replace the
affected path.

These mechanisms support decoding, resolution auditing, metadata,
and rendering. They are not a public generic conversion or cloning API.

## Safe failures

Source failures use stable `source.*` codes and YAML failures use stable `yaml.*`
codes declared by the root package. Diagnostics contain safe classifications,
semantic paths when applicable, and available source coordinates. Raw parser,
conversion, reader-error, and input text is not forwarded. Source labels and
diagnostic messages are application-visible and therefore must never contain
secrets; terminal control characters are escaped by `WriteDiagnostics`.