ConfigKit/docs/roadmap/remediation.md

# Sequence Implementation Review Remediation

Status: findings 1–17 are complete. The algorithm audit below identifies six
additional open findings, numbered 18–23. This record introduces no new release
gate.

Reviewed commit: `0be7cdb00eff576fda6b30c0600b2134d4e2e086`.
Implementation range: `bc6b977..0be7cdb`.
Authority: the sequence roadmap and implementation plan preserved at
`d0a605a:docs/roadmap/sequences.md` and
`d0a605a:docs/roadmap/implementation.md`,
[ADR 0017](../adr/0017-collection-sequence-policy.md), and
[ADR 0018](../adr/0018-cli-after-structural-resolution.md).

The public facade, shared identity/composition machinery, scoped policies,
identified-record inheritance, post-decode CLI phase, and final registered
identity validation are present. The findings below concern remaining
correctness, contract-alignment, documentation, and scaling gaps.

Validation on the reviewed commit:

- `./scripts/check.sh` passed with Go 1.26.10 on Linux.
- `go test -race ./...` passed.
- Public-API probes outside the repository reproduced findings 1–9. The probes
  used labeled byte sources, defaults, and an in-memory filesystem; they did not
  change implementation files or use external services.
- Finding 10 follows directly from the maintained text; finding 11 follows from
  the prefix-scan loop, not a timing comparison or performance threshold.

Validation after the initial remediation: `./scripts/check.sh` and
`go test -race ./...` passed with Go 1.26.10 on Linux. The original public-API
probes for findings 1–9 now demonstrate the required corrected behavior.

Validation after integration remediation: the everyday check, full race suite,
and local documentation/public-boundary checks passed with the same toolchain.
The original and follow-up public-API probes demonstrate corrected behavior for
findings 1–9 and 12–15, including unrelated underlay payload at an old version
index. The accumulated-copy correction is verified by code inspection and
ownership regressions; the contract corrections are verified against the
implemented operations and replay. Each additional finding has its own completion
record and commit.

P1 denotes the highest-priority correctness issue here; P2 denotes normal-priority
work needed to complete the selected behavior. Preserve the agreed scope while
fixing these issues. In particular, do not resolve failures by restricting valid
identity fields, requiring repeated declarations, changing merge semantics, or
adding limits.

## 1. Preserve Retained Explicit Values In Nested CLI Mapping Assignments — P1

Status: [x] Complete.

Completion: override metadata now reads the actual target subtree through one
shared lookup. Ordinary/layered regressions preserve retained values, including
zero and false, their explicit origins, and absence-based resolver behavior.
Verified `go test ./internal/engine -run 'Override|CLI' -count=1` and
`./scripts/check.sh`.

Owner: [`applyDecodedOverrides`](../../internal/engine/override_decode.go),
especially lines 59–60 and the presence/origin update helpers.

The helpers receive the complete configuration projection as `existing`, but
interpret it as the subtree at the override path. A partial nested mapping
therefore treats retained input children as absent and drops their presence and
origins. `fillDefaultOrigins` subsequently labels those retained values as defaults.

Reproduction: load the following into ordinary nested Go structs:

```yaml
settings:
  nested:
    a: 1
    b: 2
```

Apply an override at `[settings]` with value
`map[string]any{"nested": map[string]any{"a": 3}}`. The typed value of `b`
initially remains 2, but `Metadata.Lookup([settings,nested,b])` reports
`Explicit: false` and a default origin. Calling `Default` for that field in
application resolution then changes the explicitly supplied 2 to the fallback
value. This violates presence-based defaulting, not merely presentation.

Required remediation: use the actual existing target subtree when updating
presence/origins, preserving retained descendants and removing genuinely replaced
ones. Keep override batches detached and preserve the selected post-decode order.
Add regressions for nested partial mappings, retained zero/false values, metadata,
and an absence-based resolver default that must not overwrite retained input.

## 2. Read Upsert Origins At Input Paths And Write Them At Output Paths — P2

Status: [x] Complete.

Completion: sequence origin copying now separates incoming and effective paths
for precedence and inheritance. A regression covers reordered matches, appended
records, composite origins, and actual resource reads from distinct declaring
directories. Verified focused origin/inheritance/overlay/provenance tests,
`go test -race ./internal/engine`, and `./scripts/check.sh`.

Owners: [`applyOverlaySequenceEvidence`](../../internal/engine/overlay.go),
particularly lines 369 and 381, and the analogous sequence-origin copying in
[`inheritance_collection.go`](../../internal/engine/inheritance_collection.go).

The overlay helper supplies an output-index path to an origin lookup whose keys
are still incoming-index paths. When those positions differ and another incoming
record occupies the output index, metadata takes that other record's source
location. The same defect was reproduced through mapping-sibling inheritance;
its helper uses the same single-path copying pattern.

Reproduction: an underlay supplies IDs `[a,b]`; the main source declares `key: id`
and supplies `[b*,a*]`. The result is correctly `[a*,b*]`, but the value origin
for final `a*` is the source line of `b*`, and vice versa. In the probe the two
value lines were 7 and 9; final index 0 incorrectly reported line 7 instead of 9.
With separately authored records this can select a different source/context.

Required remediation: use the operation's input-to-output projection for origin
subtree reads as well as writes. Preserve composite contributing origins, not
just node-location fallbacks. Test reordered upserts and appended records with
distinct source files/locations, including nested fields, inheritance operations,
and declaring resource contexts.

## 3. Apply Tier Sequence Operations To Historical Contributions — P2

Status: [x] Complete.

Completion: per-source history now composes strict item projections with the
tier operation, including nested application prefixes and local sequence replay.
Regressions cover all three operations, multiple peers, retained records, removed
fields, source attribution, and a subsequent overlay without a false decoding
event. Verified focused history tests, `go test -race ./internal/engine`, and
`./scripts/check.sh`.

Owner: [`applyOverlayWithCollectionPolicy`](../../internal/engine/overlay.go),
lines 76–84, and existing sequence-history projection/replay.

For composed tiers, `applications` is nonempty. This branch appends the tier's
original strict-composition applications without incorporating the newly computed
precedence sequence descriptors. Replay consequently performs a whole-list
replacement or strict append rather than the actual cross-tier upsert/append.
The later decoding reconciliation conceals the mismatch by attributing a
structural combination to typed defaults/decoding.

Reproduction: lower `[a,b]` and higher `[b*,a*]` correctly yield `[a*,b*]`.
`Explain([items])` instead shows the main contribution producing `[b*,a*]`, then
a `decoding` event reordering it to `[a*,b*]` with the reason "applied retained
defaults and canonical typed decoding". No decoder performed that reordering.

Required remediation: project each retained per-source application through the
tier's selected operation while keeping peer composition distinct from tier
precedence. Do not collapse authored sources or fake decoding events. Cover
upsert, append, unmatched retained records, reordered matches, removed fields,
multiple higher-tier peers, and a subsequent selected overlay. Replay should
agree with structural assembly before real decoding reconciliation.

## 4. Reproject Nested Source Destinations After Cross-Tier Item Movement — P2

Status: [x] Complete.

Completion: layered assembly projects incoming source destinations and version
classification evidence with the same item assignments used for values/history.
Earlier contributions retain their historical coordinates, physical identities
and contexts. Main/underlay regressions cover reordered upserts, append, replace,
nested helpers, and source-local version evidence. Verified focused destination,
origin and source-version tests, focused race coverage, and `./scripts/check.sh`.

Owners: [`projectAcquiredDestinations`](../../internal/engine/imports.go) and the
sequence-operation handoff in [`assembleLayeredInput`](../../internal/engine/load_request.go).

Source destinations are projected while composing their own tier, but are not
reprojected when that tier's records move to accumulated-list positions. Nested
helper source records can therefore point at a different final record.

Reproduction: lower `items` contains `[a,b]`. Main imports a single `b` record
with `placement: sequence-item` into `[items]`; that record imports a helper
mapping supplying its value. Upsert retains `a` at index 0 and places the higher
`b` at index 1. `Sources()` nevertheless reports the surviving helper's destination
as `/items/0`, its tier-local position, rather than `/items/1`.

Required remediation: carry the sequence operation's projection through nested
source-destination/evidence updates at precedence boundaries. Keep physical
source identity and passive resource contexts unchanged, and preserve historical
meaning for displaced contributions rather than relabeling them as survivors.
Test a retained lower record followed by an imported higher record, reordered
matches, and nested helpers. Check version-evidence projection where applicable.

## 5. Honor Identity Reset On Actual Inheritance Replacement — P2

Status: [x] Complete.

Completion: inheritance resolves identity against actual child-value presence,
discarding parent YAML identity for an explicit replacement while preserving
agreement for retained values and authoritative application constraints. Added
policy-state and public-load regressions for changed/unkeyed/non-sequence
replacements, application agreement/conflict, and mode-only/append/upsert cases.
Verified focused inheritance tests and `./scripts/check.sh`.

Owners: [`inheritanceCollectionResolution` and `mergeInheritanceCollectionPolicies`](../../internal/engine/inheritance_collection.go),
especially lines 79–80 and 121.

These paths reconcile the parent's YAML identity before accounting for actual
replacement, and treat the parent's collection as retained even when the child
supplies a replacing value. This is stronger than the selected identity lifetime
and differs from source-tier replacement.

Reproduction:

```yaml
configkit:
  sequences:
    - destination: [base, items]
      key: id
    - destination: [child, items]
      key: name
      mode: replace
base:
  items:
    - id: a
      value: parent
child:
  extends: base
  items:
    - name: b
      value: child
```

Loading fails with `sequence.conflict` at `/child/items`. It should succeed:
actual whole-sequence replacement may discard the parent's YAML identity and
establish `name`. Authoritative application constraints must still apply.

Required remediation: distinguish actual replacement from a mode-only declaration
before identity agreement and transfer. Test a changed key, replacement with no
new key becoming unkeyed, permitted non-sequence replacement, and an agreeing or
conflicting application registration. Preserve rejection of identity changes
without a replacement value and during append/upsert.

## 6. Enforce Collection Operations Beneath Newly Added Ancestors — P2

Status: [x] Complete.

Completion: copying a mapping now visits nested collection operations with absent
lower values, preserving ordinary ancestor removal and whole-record semantics.
Incoming record-owned policy transfers through the same operation. Regressions
cover absent/empty/null ancestors, identity and kind failures, empty and valid
initialization, and later inheritance using an imported item's nested identity.
Corrected an older test that deferred the required error until a later tier.
Verified focused merge/collection/inheritance tests, focused ownership race
coverage, and `./scripts/check.sh`.

Owners: [`mergeRecursiveMappings`](../../internal/engine/recursive_merge.go) and
[`mergeOverlayMappingsWithCollectionPolicy`](../../internal/engine/overlay.go).

When an incoming ancestor mapping is absent below, the recursive merger clones
the entire subtree and invokes only its ancestor-level added hook. Nested
collection resolutions are not visited. This bypasses their operation rules;
the subsequent declaration-only policy update does not replace those checks.

Reproduction: combine a lower `{}` with this main tier:

```yaml
configkit:
  sequences:
    - destination: [group, items]
      mode: append
group:
  items:
    - id: a
```

Loading succeeds despite append having no identity definition. Placing the same
sequence directly at `[items]` correctly fails. The validity of an operation
must not depend on whether an enclosing mapping already existed.

Required remediation: account for applicable nested collection policies when
adding/copying an ancestor subtree, including owned-policy transfer. Test absent
and present enclosing mappings, append/upsert without identity, explicit empty
lists, invalid non-sequence contributions, and valid identified initialization.
Do not turn this into a general path-existence or shape constraint: ordinary
ancestor replacement/removal retains its documented meaning.

## 7. Validate Final Typed Identity Fields As Payload, Not Directives — P2

Status: [x] Complete.

Completion: shared identity lookup now distinguishes structural records from
effective payload/reference mappings. Final validation includes literal `extends`
fields while structural source checks still exclude actual directives. Facade
regressions cover defaults and recorded resolution, string/null components, and
duplicate rejection. Verified those regressions and existing structural identity
and inheritance coverage through `./scripts/check.sh`.

Owners: [`validateFinalSequenceIdentities`](../../internal/engine/final_identity.go)
and [`newSequenceMappingIndex`](../../internal/engine/sequence_identity.go).

The final validator reuses structural record lookup, which excludes every
`extends` entry. At this phase there is no structural processing: typed defaults
and application resolution can produce ordinary payload with that literal name.
The roadmap expressly distinguishes structural-control exclusion from final
typed-payload validation; this reuse introduces an extra restriction.

Reproduction: a defaults-only `Catalog` with `Items []map[string]any` containing
`{"extends": "application-data"}`, registered with
`SequenceIdentityAt(Keys("items"), "extends")`, fails with a missing identity
field at `/items/0/extends`, even though the effective field is present.

Required remediation: distinguish structural-payload lookup from final payload
lookup. Preserve rejection when a source tries to use an actual consumed
directive as identity. Test final fields supplied by defaults and recorded
resolution, including a null component, without adding a broader name ban.

## 8. Retain Nonfatal Main-Version Diagnostics In Layered Loads — P2

Status: [x] Complete.

Completion: layered assembly returns nonfatal main-version diagnostics on success
and alongside later failures; the shared caller stops only for errors. Regression
coverage verifies ordinary/layered warning parity, exactly one checker call, and
warnings retained with later structural or decoding errors. Verified through
`./scripts/check.sh`, including existing source-version ordering coverage.

Owner: [`assembleLayeredInput`](../../internal/engine/load_request.go), lines 143–152.

The moved default-mode version-check call returns its diagnostics only when they
contain errors. Warning-only diagnostics are discarded. Ordinary loading still
retains the same warnings.

Reproduction: a `VersionRule.Check` that calls
`validation.Warning(Keys("version"), "test warning")` and returns nil yields one
`version.warning` through ordinary `Load`, but zero diagnostics through
`LoadLayered` with the same main input and default source-version mode.

Required remediation: carry warnings from the single main-tier check into the
shared diagnostic accumulator on successful layered loads. Preserve checker
ownership, invocation count, and the selected acquisition/composition ordering.
Add ordinary/layered parity and warning-plus-later-error regressions.

## 9. Identify Conflicting Application Registrations In Diagnostics — P2

Status: [x] Complete.

Completion: definition validation retains original one-based registration
positions and names both positions in each overlap diagnostic. No input path or
new public diagnostic structure is invented. A mixed-selector regression verifies
multiple conflicts even after an earlier invalid registration is skipped.
Verified registration/facade coverage through `./scripts/check.sh`.

Owner: [`compileSequenceIdentityPolicy`](../../internal/engine/sequence_registration.go),
lines 119–128.

Incompatible dynamic overlaps produce only a generic `options.invalid` message,
an empty configuration path, and no related information. Neither registration
is identified. The roadmap requires definition-time errors to identify the
conflicting registrations without inventing a concrete input destination.

Reproduction: register `collections -> AnyMappingKey -> items` using `id`, and
`collections -> foo -> items` using `name`. Even with `NoSource`, the error is
only "overlapping sequence registrations must use the same identity fields".
With multiple registrations the caller cannot identify the conflicting pair.

Required remediation: retain original registration positions and identify both
in safe definition diagnostics, for example through positional message context.
No new public diagnostic structure is required solely for this. Test mixed exact
and dynamic registrations and multiple conflicts, without asserting complete
English wording or exposing identity values.

## 10. Remove Stale Public Contract Descriptions — P2

Status: [x] Complete.

Completion: the public definition comment now describes final registered-identity
validation and its shape/lifetime boundaries. Selected-overlay prose explicitly
separates ordinary replacement from collection policy and links to the canonical
operation/presence tables. Verified against the implementation, local-link tests,
and `git diff --check`; no runtime behavior changed.

Owners: [`Definition.SequenceIdentities`](../../lifecycle.go), lines 18–21, and
the [selected-overlay contract](../config.md#selected-overlays).

The exported definition-field comment still says registrations do not validate
identities introduced by defaults, CLI, or callbacks. Final validation is now
implemented and the registration constructors correctly document it.

The selected-overlay table and following paragraph also still unconditionally
describe sequences as atomic replacements and `[]` as clearing. A later paragraph
describes collection operations, but does not make the earlier unconditional
claims accurate: identified collections infer upsert, and `[]` then contributes
no records rather than clearing.

Required remediation: make the symbol comment describe its implemented final
guarantee and boundaries, and qualify the overlay table/paragraph for collection
policy. Link to the canonical collection tables rather than duplicating them.
Keep YAML-only enforcement distinct from registered final validation.

## 11. Remove The Quadratic Prefix Scan In Item Policy Projection — P2

Status: [x] Complete.

Completion: assigned sequence ranges now supply item offsets directly. Handoffs
without a range lazily accumulate prefix lengths, visiting each preceding
contribution once rather than rescanning every prefix. Regression coverage keeps
inline and policy-less contributions in the correct positions with complete,
partial, or absent ranges. Existing ownership/placement tests remain intact.
Verified focused item-policy tests, `./scripts/check.sh`, `go test -race ./...`,
and inspection of the remaining error-only prefix-scan callers. No new benchmark
fixture or performance threshold was needed for this bounded change.

Owner: [`projectNestedCollectionPolicies`](../../internal/engine/imports.go),
especially line 968; the scanned helper is `priorSequenceLength` in the same file.

For every item with local collection policy, the loop scans all preceding
contributions to calculate `itemIndex`, then overwrites that result when the
already-indexed sequence range is available. For N such items this performs
`0 + 1 + ... + (N-1)` contribution visits, excluding path lookup costs, despite
the common case already having the exact assigned index. This is introduced
quadratic work in the feature's independently authored-record path.

Required remediation: use the existing assigned range first. If a genuine
fallback is necessary, maintain per-destination prefix lengths in one pass
rather than rescanning every prefix. Preserve scope and ordering semantics;
do not add limits, pooling, or a new public cache. Keep ownership/placement
regressions and add a representative many-item/local-policy benchmark only if
needed to exercise the path. A wall-clock threshold is not required.

## Integration Audit Baseline And Scope

Reviewed commit: `8d40983271de5239312eb0a2896737be88c5abdc`.

This follow-up reviews the interaction of collection policy, identified-record
inheritance, dynamic registrations, and post-decode CLI assignment with existing
source placement, source-version checking, typed decoding, inspection, resource
contexts, and ownership boundaries. The accepted ADRs and current configuration
contract remain the authority; the findings require integration corrections,
not additional consumer restrictions or new feature choices.

Validation:

- `./scripts/check.sh` and `go test -race ./...` pass with Go 1.26.10 on Linux.
- Isolated public-API probes outside the repository reproduce findings 12–15.
  They use labeled byte sources and an in-memory filesystem, with no external
  services, real credentials, or repository implementation changes.
- Finding 16 follows from the per-item deep-copy loop; no timing claim or
  performance threshold is needed. Finding 17 follows from maintained contract
  text that contradicts the implemented collection operations.
- Positive comparisons include sequence-valued identity replacement, ancestor
  retirement through source tiers, quoted versions after item movement, and an
  append overlay that preserves the main record's version.

## 12. Complete Identity Retirement Across Replacement Shapes — P2

Status: [x] Complete.

Completion: source composition and inheritance now prepare eligible lower policy
from actual replacement paths before agreement. Removed identity and record state
cannot constrain a new pending key or reappear in later inheritance. Shared mode
selection preserves explicit-mode persistence and authoritative registrations.
Regressions cover null/scalar/mapping/list replacement, ancestor removal with a
new key, reintroduction, and retained/application-constrained conflicts. Verified
focused retirement/inheritance tests, engine ownership race coverage, and
`./scripts/check.sh`.

Owners: [`resolveStrictCollectionPolicies`](../../internal/engine/compose.go),
especially its `sequencePresent` reset condition;
[`mergeInheritanceCollectionPolicies`](../../internal/engine/inheritance_collection.go);
and ancestor-replacement handling in
[`applyInheritanceScopeTemplates`](../../internal/engine/inheritance_schedule.go).

Policy agreement and retirement still treat some actual replacements as retained
collections. This has three observable effects under the same selected lifetime
rule:

1. Source precedence accepts a replacement sequence with a new YAML key, but
   rejects the equivalent replacement by null or a scalar. For example, a lower
   tier declaring `key: id` for `[items]` with `items: [{id: old}]`, followed by
   a main tier declaring `key: name, mode: replace` with `items: null`, fails with
   `sequence.conflict` at `/items`. Replacing null with `42` also fails; replacing
   it with `[{name: new}]` succeeds. Null/scalar replacement should discard the
   lower identity and retain the newly declared key as pending policy, subject
   to the application's schema and constraints.
2. Removing an ancestor while declaring a new pending key for its removed
   descendant fails before the replacement occurs. A lower
   `group: {items: [{id: old}]}` with key `id`, followed by `group: null` and a
   root declaration of key `name` for `[group, items]`, fails at `/group/items`.
   The analogous `base`/`child` inheritance case fails at `/child/group/items`.
   The old collection is being removed rather than retained.
3. Inheritance can resurrect removed identity even without a conflicting new
   declaration. This input fails with a missing `id` at
   `/grandchild/group/items/0/id`:

   ```yaml
   configkit:
     sequences: [{destination: [base, group, items], key: id}]
   base: {group: {items: [{id: old}]}}
   child:
     extends: base
     group: null
   grandchild:
     extends: child
     group: {items: [{name: new}]}
   ```

   The same remove/reintroduce operations through source tiers succeed. The
   inherited descendant identity remains in policy state after its enclosing
   mapping is removed, so the grandchild incorrectly receives the old key.

Required remediation: account for actual value and ancestor replacement before
deciding which lower identity participates in agreement, and retire removed
descendant/record-owned state during inheritance as during source precedence.
Retained inheritance templates must not reestablish discarded identity later.
Preserve authoritative application constraints, independently owned mode
persistence, pending declarations, and rejection of key changes when values are
genuinely retained. Cover direct non-sequence replacement, ancestor removal with
a new pending key, and removal followed by reintroduction across both contexts.
Do not require repeated keys or narrow permitted null/scalar shapes to avoid
these cases.

## 13. Keep Metadata History In Post-Decode CLI Order — P2

Status: [x] Complete.

Completion: metadata history is assembled from the pre-CLI source view, with CLI
events appended separately; inheritance is projected before CLI and resolution.
Effective fields still describe post-CLI values. Regressions verify chronological
origins and resolver dependencies for ordinary/layered mapping and record
inheritance, including equal-valued assignments. Verified history/override
coverage and `./scripts/check.sh`.

Owners: [`historyWithInheritance`](../../internal/engine/inheritance_history.go)
and the source/CLI metadata handoff in
[`preparePostDecodeOverrideContext`](../../internal/engine/override_context.go).

The value explanation follows the new lifecycle, but `Metadata.History` still
inserts inheritance after stored CLI changes. `historyWithInheritance` splits
only at `LayerResolution`, reflecting the former CLI-before-inheritance order.
The reconstructed source view can also append an authored source origin after
the CLI event.

Reproduction: load `base: {value: parent}` and
`child: {extends: base, value: authored}`, then override `[child, value]` with
`"cli"`. The final value and `Metadata.Lookup` correctly report CLI provenance.
`Explain` reports source → inheritance → CLI. `Metadata.History` instead reports
source → CLI → source → inheritance and ends with a source/inheritance origin.
This violates chronological metadata history and disagrees with the effective
origin and the value explanation.

Required remediation: assemble metadata history in the implemented source,
inheritance, CLI, and resolution order, preserving the real origins and
dependencies of those events. Cover ordinary/layered loading, mapping and record
inheritance, equal-valued CLI writes, and a later recorded resolver change at the
appropriate existing test owners. Do not impose a new exact event-count or
message-format contract.

## 14. Carry Main-Version Coordinates Through Sequence Movement — P2

Status: [x] Complete.

Completion: the effective version rule now follows the main declaration through
strict placement, tier combination, and inheritance without changing source-local
checker/evidence paths. Underlay and overlay ownership checks use assigned record
positions, including whole-record removal; integer conversion remains confined
to the actual main header. Regressions cover mounted/indexed paths, appended and
matched records, integer/string parity, unrelated version payload, imported
underlay claims, and inheritance. Corrected an older ownership fixture to target
the actual matching main record. Verified source-version regressions,
`./scripts/check.sh`, and `go test -race ./internal/engine`.

Owners: [`effectiveSourceVersionRuleForPlacement`](../../internal/engine/source_version.go),
the main-tier projection handoff in
[`assembleLayeredInput`](../../internal/engine/load_request.go), and use of
the effective rule in [`executeLoadRequest`](../../internal/engine/load.go).

Source destinations and version-evidence sensitivity paths follow item movement,
but the rule used for integer-version decoding retains the original item index.
Consequently an accepted integer header can fail decoding after an otherwise
valid upsert placement.

Reproduction with the public API:

- Main input is `id: main` and `version: 4`, placed as one sequence item at
  `Keys("items")`.
- A lower input contributes the sequence item `id: lower` at the same destination.
- Register identity `id` for `[items]`, select `SourceVersionModeIfPresent`, and
  use `VersionRule.Path = Keys("version")` with an accepting checker.
- Decode into records with `ID string` and `Version string`.

The checker accepts `4` exactly once. Upsert places the main record at index 1,
but decoding fails with `decode.type` at `/items/1/version`. The integer-to-string
exception is still tied to `/items/0/version`. Quoting the identical version as
`'4'` succeeds. A source-local indexed header path, without item placement,
reproduces the same problem when its record moves during upsert.

Required remediation: carry the main declaration's effective coordinates through
the existing item projections into decoding and audit the associated ownership
protection handoffs for the same stale-coordinate assumption. Preserve
source-local checker/evidence paths, checker invocation count, main-only
ownership, and the narrow integer-to-string exception. Cover moved existing and
newly appended main records, integer/string parity, and protection of the actual
main header after movement. Do not prohibit supported indexed paths or mounted
main records as a workaround.

## 15. Exclude Parent Identity From Inherited History Payload — P2

Status: [x] Complete.

Completion: retained record-inheritance operations now exclude parent identity
from their payload and origin projection. Replay preserves the child's own
identity before taking field, record, sequence, or root snapshots, without
inventing inheritance events for retained identity fields. Regressions cover
single and compound identities, explicit child provenance, ordinary inherited
payload, and the parent's original history. Verified inheritance/inspection
coverage, `./scripts/check.sh`, and `go test -race ./internal/engine`.

Owners: inheritance operation construction in
[`applyInheritanceScopeTemplates`](../../internal/engine/inheritance_schedule.go),
[`appendInheritanceExplanation`](../../internal/engine/inheritance_explain.go),
and transferred-origin projection in
[`inheritance_history.go`](../../internal/engine/inheritance_history.go).

Identified-record value construction excludes the parent's identity fields, but
the retained history operation still supplies the complete parent record and
its identity origins to the older mapping-inheritance replay machinery.

Reproduction:

```yaml
configkit:
  sequences: [{destination: [items], key: id}]
items:
  - id: base
    value: parent
  - id: child
    extends: {id: base}
```

The effective child ID and origin are correct. However,
`Explain(Keys("items").Index(1).Key("id"))` reports `child` → `base` → `child`,
attributing the intermediate value to an inheritance operation and the parent's
ID location. `Metadata.History` likewise claims an inherited parent-ID
contribution. That identity is expressly excluded from inherited payload; it was
never an applied value at the child path.

Required remediation: make history replay and transferred-origin projection use
the same identity exclusion as value construction. Preserve the child's authored
identity and presence, ordinary inherited payload history, and the parent record's
own source history at its original path. Cover compound identities and identity
fields alongside inherited payload at both record and ancestor queries; keep the
existing inspection API and sensitivity behavior.

## 16. Avoid Deep-Copying Accumulated Item Policies Per Record — P2

Status: [x] Complete.

Completion: item projection now takes one owned policy copy per destination,
updates that local state incrementally, and stores it once after projection.
Incoming scopes still cross detached ownership boundaries. Inspection confirms
that accumulated record scopes are no longer deep-copied per item; no timing
threshold or benchmark campaign was needed. Existing stable-index, sparse-range,
fallback-offset, and ownership coverage passes, with a new regression for
interleaved destinations, distinct record owners, and sibling/cross-load mutation
independence. Verified focused item-policy tests, `./scripts/check.sh`, and
`go test -race ./...`.

Owners: [`projectNestedCollectionPolicies`](../../internal/engine/imports.go)
and its use of copying `collectionPolicyScope.at` / `set` helpers in
[`collection_policy_state.go`](../../internal/engine/collection_policy_state.go).

The prefix-scan correction in finding 11 is present, but a separate quadratic
cost remains on the same independently authored-record path. For every imported
item with local policy, `at` deep-clones the destination's accumulated record
order and every retained record's nested scope. After adding one owner, `set`
deep-clones the growing policy again. With N fixed-size item-local policies at
one destination, this copies approximately `1 + 2 + ... + N` retained scopes,
even when every assigned sequence range is available.

This follows directly from the loop and `cloneCollectionPolicy`; no busy-host
timing comparison is required. It is an integration cost from reusing detached
read/write helpers within private incremental construction, rather than from
the required final ownership copies.

Required remediation: construct each destination's owned record-policy state
without repeatedly cloning all earlier records, retaining necessary copies at
ownership boundaries. Preserve record ordering, distinct owners, sparse assigned
positions, fallback offsets, and cross-load detachment. Reuse the existing item
policy and ownership tests; add allocation/scaling evidence only where useful.
Do not introduce public caches, pooling, new caps, or performance thresholds.

## 17. Reconcile Legacy Sequence-History Contract Text — P2

Status: [x] Complete.

Completion: metadata and explanation prose now describe history through the
selected collection operation, assigned item positions, and displaced payload.
The text distinguishes sequence snapshots and atomic diffs from append/upsert
combination and links to the canonical operation tables. Whole-record upsert,
existing queries, and sensitivity guarantees are unchanged. Verified against
replay/combination code, local documentation-link tests, and `git diff --check`;
the unchanged runtime is covered by the preceding everyday and full race checks.

Owner: the metadata/provenance and explanation sections of
[`docs/config.md`](../config.md#layered-metadata).

The metadata section still says that a later tier supplying a sequence records
replacement at the sequence path and that history does not imply identity-based
merging across tiers. That describes the former unconditional replacement model,
not the implemented append/upsert history or the collection operation tables in
the same document. The explanation paragraph about atomic sequences also needs
review so readers can distinguish snapshot/diff granularity from collection
combination semantics.

Required remediation: describe history for the selected collection operation and
link to the canonical operation tables. Preserve the existing atomic sequence
snapshot/diff behavior and whole-record upsert semantics; correcting the prose
does not authorize recursive record-field merging, a new query API, or broader
history guarantees.

## Algorithm Audit Baseline And Scope

Reviewed commit: `c73daf891d0f29c96f7026bdabcccb571f44a751`.
Feature baseline: `4f941f8ba4c9a9ed0afa87b07fe6ba29f131d1c6`.

This audit examines the sequence identity/operation, scoped-policy, inheritance,
inspection, and post-decode CLI algorithms, including the completed corrections
above. The intervening discovery roadmap commits add no implementation to audit.
The accepted scheduler, deterministic ordering, exact numeric normalization,
selector trie, and deliberately pairwise registration-overlap validation remain
selected design choices; they are not findings here.

Evidence and validation:

- Code inspection establishes the complexity of findings 19–23 on bounded-depth
  inputs with linear-sized results. Required result/history copies and genuinely
  growing inheritance output are excluded from the alleged excess work. Bounds
  describe internal operations, not new public performance guarantees.
- An isolated public-API probe reproduces finding 18 using an in-memory
  filesystem, without external services or real credentials.
- Temporary tests supplied through a Go overlay, outside the repository, checked
  all 768 combinations of append/upsert/replace over the 16 unique-ID lists drawn
  from three IDs. Stable ordering, duplicate rejection, and whole-record payload
  replacement agreed with an independent reference calculation.
- A structural overlay test confirmed that 4, 8, and 16 one-item contributions
  retain 10, 36, and 136 reverse-index cells respectively in `indexSequenceRanges`.
  These are representation counts, not timing or real-world benchmark results.
- `go test ./internal/engine -run 'Sequence|Collection|Inheritance|Override|SourceVersion|Metadata|History' -count=1`
  passed with Go 1.26.10 on Linux. Passing existing tests does not cover the
  incorrect selector scope or establish efficient growth.

The following are focused corrections within the selected model. This review
changes only the remediation record.

## 18. Match Application Registrations In Effective Item Coordinates — P2

Status: [x] Complete.

Completion: pending item/helper scopes now compose their local YAML policy without
reinterpreting global mapping-only application selectors below a sequence index.
The enclosing composer still enforces the actual outer collection constraint.
Updated the old unit test to exercise that boundary and added ordinary/layered
facade regressions for exact/dynamic unused selectors, nested helpers, and missing
outer identities. Verified the facade regressions and `./scripts/check.sh`.

Owner: `composePendingItems` in
[`imports.go`](../../internal/engine/imports.go), especially its local
`composeDocumentsWithCollectionPolicy` call; the current
`TestPendingItemCompositionAppliesApplicationIdentityConstraints` in
[`item_policy_test.go`](../../internal/engine/item_policy_test.go) also needs review.

Pending item/helper documents are composed in item-local coordinates while the
unchanged global application selector policy is supplied to their composer.
A registration for root `[prompts]` therefore also constrains an unrelated
`[items, 0, prompts]` collection before placement. This contradicts complete,
effective-destination matching and makes an otherwise unused registration reject
valid configuration. Application selectors do not traverse sequence indexes.

Reproduction: main imports `item.yml` with destination `[items]` and placement
`sequence-item`. The item contains `id: outer` and
`prompts: [{text: payload}]`. Loading succeeds without a registration; adding
only `SequenceIdentityAt(Keys("prompts"), "id")` fails with `compose.invalid`
at `/items/0/prompts/0/id`, although no root `prompts` collection exists.

Required remediation: separate item-local YAML policy from application matching
in effective coordinates, including nested helper scopes. Preserve applicable
outer collection constraints without reinterpreting global selectors as local
ones. Add ordinary/layered regressions for unused root and dynamic selectors
alongside imported item-local collections. Correct tests that pin the accidental
scope; do not change selector meaning or require consumers to remove valid rules.

## 19. Avoid Dense Global-Position Storage For Sparse Item Projections — P2

Status: [x] Complete.

Completion: contiguous composition ranges and their origin contributions retain
arithmetic offset lookup. Scattered projections use independently cloned sparse
maps in both directions; unprepared range queries no longer allocate indexes.
Audited the shared projection/history callers for batch preparation. Added
structural coverage for many one-item ranges, widely separated assignments, and
absent coordinates. Existing upsert, history, version, and ownership tests pass,
along with `./scripts/check.sh` and `go test -race ./internal/engine`.

Owners: `newSequenceItemProjectionIndex` in
[`sequence_projection.go`](../../internal/engine/sequence_projection.go),
`indexSequenceRange` / `projectedCompositionOrigins` in
[`compose.go`](../../internal/engine/compose.go), and contribution indexing in
[`history.go`](../../internal/engine/history.go).

The reverse projection slice is sized to `maxOutput + 1`, even for one incoming
item assigned to a late assembled position. `indexSequenceRange` materializes
these indexes for ordinary contiguous append ranges too. With N one-item sources
contributing to one sequence, reverse slices retain `1 + 2 + ... + N` cells.
Initialization time and simultaneously retained auxiliary storage are therefore
quadratic despite only N actual mappings. Origin/history projection constructs
the same representation through additional callers. Finding 16's policy-copy
correction does not remove this independent cost.

Required remediation: retain arithmetic start/end lookup for contiguous ranges
and use storage proportional to actual mappings for scattered positions, such as
a sparse reverse map. Audit shared callers and avoid rebuilding equivalent indexes
inside item queries. Preserve bidirectional lookup, absent positions, upsert
replacement evidence, ordering, and detached results. Verify the representation
on many one-item sources and noncontiguous upserts; add no timing threshold or
collection-size limit.

## 20. Remove Full Metadata Scans Per CLI Replacement Leaf — P2

Status: [x] Complete.

Completion: presence and origin removal now visits only the known replaced
pre-CLI subtree. Partial mappings still retain omitted fields and their metadata.
Added broad-mapping coverage with escaped keys, explicit zero/false values,
empty-list replacement, null pointers, retained siblings, and deleted descendant
metadata. Existing nested CLI and typed-ownership regressions and
`./scripts/check.sh` pass.

Owners: `updateOverridePresence`, `deletePresencePathSubtree`, and
`updateOverrideOrigins` in
[`override_decode.go`](../../internal/engine/override_decode.go).

Every replaced scalar or subtree scans all presence keys, then all origin keys,
to find matching prefixes. A single partial-mapping override updating N scalar
children in an N-field configuration performs quadratic map-entry visits. This
does not require N CLI sources or growing historical output: the same effective
maps remain approximately N entries throughout the update.

Required remediation: delete affected entries by walking the already known
existing subtree, or use one batch-local prefix index where needed. Retain
necessary copies at ownership boundaries without rescanning unrelated paths for
each leaf. Preserve retained explicit zero/false values, removal of replaced
descendants, null/kind changes, origin attribution, and detached loads. Use the
existing nested CLI and ownership regressions plus focused broad-mapping coverage;
do not change override semantics or impose batch limits.

## 21. Localize Inheritance Policy Work To The Relevant Scope — P2

Status: [x] Complete.

Completion: inheritance now keeps a per-load semantic path index, including
separate indexes for record-owned scopes. Walker handoff and descendant
retirement touch the changed scope; mapping and identified-record preparation
reuse indexes across sibling queries. Only relevant policies are cloned into
independent templates. Added wide independent YAML groups, pending-policy
replacement, literal-key/index distinctions, and ownership regressions. Existing
nested item/inheritance, mode-lifetime, and application-constraint coverage,
`./scripts/check.sh`, and `go test -race ./internal/engine` pass.

Owners: `inheritancePolicyScopeAtPath` in
[`inheritance.go`](../../internal/engine/inheritance.go), graph preparation in
[`inheritance_schedule.go`](../../internal/engine/inheritance_schedule.go),
and policy handoff/retirement in
[`inheritance_walk.go`](../../internal/engine/inheritance_walk.go).

Scope localization scans every entry in the supplied policy table for each
definition. Mapping graph preparation also clones the full supplied table, and
the walker passes its global table to each independent scope and scans that table
again during replacement. For S scopes or siblings and P unrelated collection
policies, these paths perform repeated S-by-P work and copies.

For example, N independent mapping groups can each contain a constant-sized
`base`/`child extends base` pair and one YAML collection declaration beneath its
base. Input, expanded values, and relationship evidence remain linear in N, but
each group's graph/localization visits or clones the other groups' policies.
This is separate from necessary expansion along a genuinely growing parent chain
and from the selected dependency scheduler.

Required remediation: partition or index policies by semantic scope once and
carry the relevant state through traversal, including efficient descendant
retirement as inheritance introduces or removes destinations. Clone owned
in-scope state rather than unrelated global state. Preserve record owners,
pending declarations, mode lifetime, destination-local registration matching,
and independent templates. Verify wide independent groups and nested item scopes;
do not replace the agreed scheduling algorithm or add global caches.

## 22. Index Fallback Location Lookup During Effective Projection — P2

Status: [x] Complete.

Completion: effective location projection lazily builds one read-only fallback
document lookup and shares it across field and entry visits. The ordinary
nearest-location helper retains its existing behavior. Added broad mixed
authored/default mappings, ancestor fallback, literal key/index mismatch,
synthetic CLI coordinates, and detached-projection coverage. Existing secret
and validation diagnostics and `./scripts/check.sh` pass.

Owner: `documentWithMetadataLocations` in
[`override_context.go`](../../internal/engine/override_context.go), using
`nearestDocumentLocation` in
[`validation.go`](../../internal/engine/validation.go).

The new effective-projection walk calls the unindexed nearest-location helper
for each field lacking a real origin location, often once for its mapping entry
and again for its value. Each call starts at the fallback root and linearly scans
mapping entries. With N authored root fields and N different default-only root
fields, each default's failed lookup scans all N authored keys: quadratic work
for linear-sized input and output. The shared lifecycle calls this path even
without CLI assignments or a resolver.

Required remediation: share an indexed fallback lookup or traverse corresponding
containers together while carrying the nearest available location. Existing
`documentLookup` / mapping-index machinery can support bounded-depth lookups
without repeated wide scans. Preserve nearest-ancestor fallback, source IDs,
synthetic locations, secret-reference diagnostics, and publication ownership.
Exercise mixed authored/default-only broad mappings; do not weaken attribution
or skip needed location attachment to avoid the work.

## 23. Index Repeated Collection-Policy Document Lookups — P2

Status: [ ] Open.

Owners: `mergeOverlayMappingsWithCollectionPolicy` in
[`overlay.go`](../../internal/engine/overlay.go) and
`mergeInheritanceCollectionPolicies` in
[`inheritance_collection.go`](../../internal/engine/inheritance_collection.go).

These routines loop over collection policies/declarations and call plain
`documentAtPath` for each destination. That helper linearly searches every
mapping on the path. N independent collections below N sibling mapping keys
therefore cause quadratic mapping-entry searches when checking which lower
policies the incoming tier touches. The inheritance parent/child lookups and
retained-value declaration checks have the same pattern. YAML declarations alone
can establish these policies; the deliberately pairwise application-registration
overlap check is not involved.

Required remediation: reuse a document lookup built once for each stable input
tree, or use the already collected claims where their scope is sufficient.
Index a mutated result only after its value-changing phase, rather than retaining
a stale lookup. Preserve absent destinations, non-sequence pending declarations,
literal key/index distinctions, retirement, and error locations. Verify broad
independent collections across precedence and inheritance; no new public cache,
shape requirement, or performance gate is needed.