promptkit/docs/roadmap/audit-remediation.md

# Audit Remediation Roadmap

## Purpose And Status

Prepare PromptKit's audited implementation for a focused round of correctness,
configuration simplification, resource-use improvements, and internal cleanup.
This document defines the target and work boundaries. The companion
[implementation plan](implementation.md) owns staged steps and validation;
neither document describes completed remediation yet.

Baseline: repository commit `bd3dfa9`, with the
[audit catalog](audit/catalog.md) covering 41 source findings consolidated into
34 remediation entries. All six catalog decisions have been accepted; their
implementation remains pending. The planning assessment was completed at
`3ef2a9e` against the unchanged production baseline, with evidence gathering
starting at `21fea46`. Its recommendations, measurements, and subsequent review
refinements are consolidated here; the separate assessment document is retired.
The catalog and supporting audits own finding evidence and decision rationale.
This roadmap owns the selected end state, work boundaries, assessment evidence,
and workstream status. Earlier remediation suggestions in the source audits do
not override accepted decisions.

All workstreams below are pending implementation; their planning assessment is
complete. No finding is proposed for deferral. Concrete designs below are the
selected planning recommendations, not claims of tested replacement behavior.
Private helper names are descriptive unless identifying an existing declaration.
Catalog numbers are traceability identifiers, not a required implementation
order or 34 separate delivery stages. The scope
contains confirmed defects, documentation/test gaps, and maintainability work;
those categories do not have equal urgency or automatically become release gates.

## Targeted End State

PromptKit remains an application-neutral Go library with the existing public
facade and ConfigKit composition boundary. The intended changes are:

- Programmatic backend/profile sources hand supported values to ConfigKit
  directly, removing PromptKit's recursive compatibility conversion.
- Schemas supplied for validation and native structured output are self-contained
  documents, while in-document references remain supported.
- Artifact-reader ownership, public errors, and smoke-test outcomes accurately
  express the owning operation and collected evidence.
- Authentication response consumption is bounded by the selected internal
  allowance, without treating overflow as rejected credentials.
- Identified unnecessary retention, repeated traversal, and temporary allocation
  are reduced with proportionate changes that preserve established behavior.
- Production execution and its tests use the real generation coordinator;
  obsolete adapters, state, helpers, and duplicate test expectations are removed.
- Current documentation describes actual behavior, and deliberate compatibility
  changes are explained where consumers encounter them.

The [resolved decisions](audit/catalog.md#resolved-decisions) are authoritative
for the selected changes. They do not authorize stronger numerical precision,
provider compatibility, scheduling, hash stability, or performance guarantees.

## Scope Boundaries

Keep the work within the [architecture policy](../policy/architecture.md) and
[configuration format boundary](../formats.md). ConfigKit owns source composition
and its raw-source diagnostics; PromptKit validates its final effective values.

This round does not include a schema bundler, a new local-only schema mode, an
upstream ConfigKit extension, a numeric conversion subsystem, new inference
limits or application deadlines, or a general reflection/validation framework.
[Schema bundling](future.md#schema-bundling) remains a future idea. Runtime
provider overrides retain their distinct supported-value and copying behavior.

Refactoring does not authorize expanding recognized template expressions,
introducing additional retries, adding pagination or runtime inventory caches,
or redesigning the capacity manager. Tests should protect established semantics,
not turn incidental implementation choices into new requirements. Existing
privacy and ownership boundaries remain applicable throughout.

## Workstreams

### A. Maintainer Toolchain Baseline

**Catalog:** 1. **Decision:** 5.

**Current gap:** The audited local toolchain predates the networking fixes
identified by the security audit. Toolchain selection and consumer compatibility
must remain separate concerns.

**Assessed baseline:** Default `/snap/bin/go` reports 1.27.0. The assessment
found an executable Go 1.27.3 installation at
`/home/eric/go/pkg/mod/golang.org/toolchain@v0.0.1-go1.27.3.linux-amd64/bin/go`;
other inspected installations were older. No installation is needed for the
assessed environment. ConfigKit 1.1.0, jsonschema/v6 6.0.2, and ReleaseKit 0.9.0
remain pinned. The focused version/applicability check found that 1.27.3 excludes
the affected ranges of GO-2026-6611 and GO-2026-6809 identified in the audit.
This was not a new full scan or an assessment of consumer binaries.

**Selected work:** Put the newest available local Go binary directory first on
`PATH` and use `GOTOOLCHAIN=local` for ordinary builds, tests, and release
validation. The existing ReleaseKit launcher already uses local selection.
Use the assessed 1.27.3 installation unless a newer local version is available
when implementation begins. Clarify this in contributor guidance. Assessment
commands also used `GOWORK=off` for isolation; that is not a new restriction on
consumer workspaces.

Leave `go 1.27.0` unchanged for the assessed toolchain. Follow the
[release procedure](../release.md#prepare-the-candidate) and
[consumer-minimum policy](../policy/release.md#consumer-minimum-and-release-toolchain)
for later series changes and existing exceptions. Selecting a newer validation
patch does not itself raise the consumer minimum. Continuing with default
1.27.0 is unnecessary given the installed fixed patch.

**Completion evidence:** Record the actual validation toolchain and retain the
focused reassessment above rather than repeating it without changed evidence.
Local installation work is needed only if the selected installation becomes
unavailable or new evidence calls for it. No recurring scan gate or consumer
advisory about generally applicable Go vulnerabilities is introduced.

### B. Configuration And Schema Simplification

**Catalog:** 3, 5, 6, 15, 21. **Decisions:** 2 and 4.

**Current gap:** Convenience-source conversion broadens ConfigKit's accepted
input forms by recursively rebuilding them, causing silent numeric changes and
potentially exponential expansion. Schema validation loads external resources
locally but sends only the root document to the provider. Compilation is
repeated across direct/prepared and host-file/`fs.FS` paths.

**Selected source handoff:** Assign `config.ExtraParams` directly into the
ordinary fields map in `ProfileSource` and `BackendSource`, then call
`configkit.Mapping` as now. Optional typed fields and the `[]any` definition
wrapper already use supported forms. The recursive compatibility adapter in
`provider_values.go` is retired. ConfigKit's immediate snapshot and deferred
unsupported/cyclic diagnostics make an extra shallow copy unnecessary. Do not
add a replacement converter, precision checker, memoization, or raw-source
validation pass.

**Selected schema design:** Retain source-specific root loading, then share
`compileSchemaDocument(ctx, resourceURL, document)` for compilation. Register
the decoded root with `Compiler.AddResource`, retain the existing Draft2020
selection, and use a context-aware loader that rejects requests for external
resources with an actionable error and no filesystem/network I/O. The compiler
continues to own URI resolution, embedded IDs, anchors, and bundled metaschemas.
Send the self-contained root document to the provider unchanged in meaning.

The dependency probe found that unused `$defs` entries are not necessarily
compiled with the root: a denying loader alone misses their external references.
Enumerate schema locations with a small explicit table of schema-bearing
keyword positions supported by the pinned dependency, and compile those
locations through the same compiler/cache. Document the table's dependency
relationship. Keep this a location traversal, not a second resolver or general
schema implementation. A lexical ban on non-fragment references would reject
valid embedded IDs; traversing arbitrary objects would mistake annotation data
for schemas. Neither is selected.

Cover escaped JSON-pointer locations, annotation/example objects, embedded IDs,
anchors, recursion, and dynamic references. Check embedded-resource dialect
handling explicitly rather than assuming the root dialect governs every nested
resource. Preserve existing supported dialect behavior; do not use this walk to
introduce broader dialect support or new restrictions. The probe recipe and
observed results are retained under [planning evidence](#planning-evidence).

Direct validation invokes shared compilation only after instance decoding;
prepared validation compiles eagerly as before. Remove external-resource loaders,
not root containment/read helpers. Change runner fields/constructor parameters
and the internal engine Options validator slot to `validate.ValidationPreparer`.
Return concrete validator pointer types from internal constructors so direct
`Validate` and preparation callers both compile. Preserve nil-validator no-op,
nil-plan error, cancellation, and frozen-plan behavior. No public facade
signature change is required.

Cover `WithSchemaFile` through public construction and preparation/validation,
including source replacement/precedence and relevant file errors. Its unused
private helper result is retired as described in F. External-resource success
fixtures become early-error cases, retaining independent root-source tests.

**Completion evidence:** Focused source-helper cases demonstrate supported-value
handoff, immediate source capture, and ConfigKit's deferred unsupported/cyclic
value diagnostics. Include formerly converted input forms to demonstrate the
migration boundary: nested built-in maps/lists, mutation after source creation,
and representative `json.Number`, named scalar/typed container, pointer, and
cyclic inputs for both helpers. Keep optional-field and construction-order
coverage; do not duplicate ConfigKit's entire limit suite. Schema cases establish external-reference rejection before provider dispatch, retained
in-document references, unused external-definition rejection, and an outbound
self-contained constraint usable without other local files. Retain independent source-path and direct/prepared coverage.
These checks establish packaging completeness, not universal provider support.

**Compatibility:** Explain caller conversion of `json.Number`, named scalars,
typed containers, arrays, and nested pointers into ConfigKit-supported forms.
For example, replace `[]string` with `[]any` containing strings; callers choose
and check numeric conversion themselves. Explain removal of external schema
loading. No arbitrary-precision promise or change to runtime request overrides
accompanies this simplification.

### C. Ownership, Errors, And Diagnostic Evidence

**Catalog:** 2, 4, 7, 8, 9, 11, 26, 27. **Decisions:** 1 and 3.

**Current gap:** The documented artifact buffer lifetime permits races; nested
errors can obscure their owning operation; response read failures can be
misclassified; some pre-dispatch smoke failures produce a completed report;
and successful authentication responses are read without a byte bound.

**Target properties:**

- Artifact GoDoc, adapters, and tests align with successful `Read` transferring
   the artifact and mutable body storage to PromptKit. Readers stop mutating
   that storage, including retained aliases. Keep the interface signature and
   existing defensive copy initially. Readers may reuse private buffers when
   they return an independent copy; no release protocol is introduced.
- Preserve the owning public error category independently of nested causes,
   retaining appropriate `errors.Is`/`errors.As` access. Do not require category
   exclusivity: retained causes can make more than one category match. Preserve
   meaningful direct-not-found versus load-failure distinctions. Use a small
   internal operation error carrying category/cause with `Unwrap`/`Is` behavior.
   Apply it at owning boundaries and select its category in facade translation,
   including before nested capacity/provider/exhaustion projections can obscure
   it. Preserve those specialized public projections and matchable child causes.
   Reordering a whole-chain sentinel switch cannot express operation ownership.
- Distinguish successful-response body I/O failures from malformed JSON at both
   initial decoding and end verification. Preserve the redacted transport
   boundary, cancellation precedence, and genuine malformed, truncated, or
   oversized-content behavior without adding retries. Record actual non-EOF
   reader errors beneath the existing limited reader. On decode/end-check
   failure, prefer observed context errors, retain size-overflow classification,
   and otherwise propagate recorded I/O failures through `requestFailedError`.
   Distinguish reader-returned errors from parser-generated syntax/type/truncation
   errors, rather than treating every `io.ErrUnexpectedEOF` alike.
- Return local inference-diagnostic construction failures as `ErrInvalidRequest`
   using explicit preparation-stage information. Preserve completed diagnostics
   for missing credentials, capacity outcomes, and unsupported/inapplicable
   operations; absence of dispatch alone does not imply an invalid request.
   Add internal `GenerationEvidence.PreparationFailed bool`, set only at the
   built-in `PrepareGeneration` failure return. Diagnostic orchestration checks
   it after caller cancellation. Injected-client errors retain their established
   semantics; no new public stage field is needed.
- Define the authentication response allowance as a named internal constant of
   1 MiB (1,048,576 bytes). Enforce against actual reads regardless of declared
   length, detect overflow, and close without unbounded draining. Report an
   explicit limit diagnostic with credential acceptance unknown and retain
   available HTTP status/timing. Add matching public/domain reason constants
   `SmokeTestReasonAuthenticationResponseLimitExceeded` with value
   `authentication_response_limit_exceeded`. For HTTP 200 recognition overflow,
   use existing `ExecutionFailureOther`, nil `CredentialsAccepted`, and a
   completed report with nil Go error unless caller cancellation applies.
   No new report field, failure enum, or public budget knob is needed.
   Read at most allowance plus one byte using `io.LimitReader`/`io.ReadAll`;
   check overflow before JSON recognition. Preserve 401 rejection without body
   consumption and other non-200 behavior independently of unused bodies.
- OpenRouter-specific authentication request/response interpretation belongs in
   the outbound adapter. Leave selection, applicability, completion, and neutral
   reporting in the use case. Preserve anonymous probing and credential/body
   privacy. Narrow smoke copying to actual consumers, removing unused request
   cloning and replacing full-report cloning used only to detach a target.
   Use an `internal/llm` authentication operation returning HTTP evidence,
   optional acceptance, overflow/unrecognized indicators, and operational error;
   the use case maps those facts to report outcomes. Move default endpoint
   derivation to an llm helper used by selection, retaining explicit overrides.
   Consume a narrow authentication capability alongside reachability HTTP,
   without reverse dependencies or a plugin registry. Directly detach the
   target/provider map; retain facade and active-backend copying.
- Exercise `RunRequest`'s privacy formatting and JSON omission boundary with
   synthetic markers in omitted variables, bodies, and execution settings, not
   IDs deliberately present in summaries. Cover `%v`, `%+v`, `%#v`, direct
   String/GoString, value/pointer forms, and direct-key JSON omission.

**Completion evidence:** Focused concurrent artifact tests follow the selected
ownership contract; they do not promise safety for readers that violate it.
Offline regressions cover composed public causes, both response read-failure
locations, construction failures with no dispatch, and authentication boundary
and overflow cases including exact limit/limit-plus-one, missing/inaccurate
lengths, closure without draining, 401 without consumption, context interruption,
and redaction. Update reason GoDoc and public conversion with the bound. Assert
diagnostic facts and privacy rather than complete incidental serialized shapes. Preserve active
backend copying, facade projection, and provider-map error behavior.

The public-boundary outcome matrix for planning is:

| Case | Owning outcome | Evidence preserved |
| --- | --- | --- |
| Client returns a child missing-prompt error | `ErrLLMGenerate` | Matchable child cause/category. |
| Reader returns that child error | `ErrArtifactLoad` | Matchable child cause/category. |
| Direct missing prompt/profile | Existing not-found error | No generic load-error substitution. |
| 2xx body I/O error at either decode stage | Generation/transport failure | Original read cause with redacted formatting. |
| Parser-invalid/truncated JSON, extra value, inference size overflow | Existing malformed-response behavior | No additional retry or parser-detail disclosure. |
| Built-in diagnostic preparation failure | `ErrInvalidRequest`, not a completed report | Zero dispatch and preparation cause. |
| Missing required credential or capacity outcome | Existing completed diagnostic | No inferred construction failure. |
| Authentication recognition overflow | Explicit limit reason; acceptance unknown | HTTP/timing facts, not secret-bearing response data. |
| Caller cancellation | Existing cancellation precedence | No publication of success over cancellation. |

### D. Resource Use And Inspection Efficiency

**Catalog:** 10, 16, 17, 18, 19, 20.

**Current gap:** The audited paths repeatedly copy requirement evidence, retain
unneeded backing storage or engine ownership, serialize unbounded diagnostic
text before truncation, reclone overlay descendants, and scan complete catalogs
for namespace-filtered listings.

**Selected remedies and completion evidence:**

- **Requirement evidence (10):** Intermediate missing-name sets are owned by
  their evaluation results. Merge the smaller set into the larger, transfer nil
  sets directly, and combine flags as now. Operands subsequently read values and
  known flags, not independent evidence snapshots. This avoids repeated growing
  copies without a global accumulator that could mix branch-local evidence.
  Preserve union/flag outcomes and conservative inspection coverage with
  representative left-heavy cases and independent renderer comparisons.
- **Prepared handles (16):** Replace `owner *Runner` with a pointer to an immutable,
  non-zero-sized private identity object allocated once per runner constructor.
  Keep the token after claim/discard so foreign-owner checks retain their order
  without retaining the runner graph or adding mutable owner-field races.
  Retain mutex-protected state transitions/payload clearing and detached details.
  Cover foreign owners, namespace views, concurrent claim/discard, one-shot use,
  and details after use/discard. A GC-timing assertion is unnecessary.
- **Repair feedback (17):** Attempt bounded construction of the complete-list
  result; if it cannot fit, construct the omission form with existing order and
  reservation rules. Normalize only potentially useful prefixes, preserving
  `strings.ToValidUTF8` behavior for invalid-byte runs and rune-safe truncation.
  Use standard `json.Marshal` on bounded normalized strings and existing prefix
  selection rather than a custom escaper. A limited writer around `json.Encoder`
  does not suffice if encoding first allocates the whole original string.
  Preserve the 64 KiB output rule, omission marker, valid JSON, HTML/control
  escaping, Unicode, complete stored diagnostics, and previous candidate.
  Cover fitting/exact-boundary/oversized/many-entry/escaped/invalid inputs.
- **Session IDs (18):** Apply `strings.Clone` to the validated nonempty normalized
  value before returning from `NormalizeSessionID`. Preserve trimming, Unicode,
  accepted content, and wire/hash values without adding a raw-input limit.
- **Provider overlays (19):** Clone the lower map once, then recurse into owned
  descendants for overlapping built-in maps and copy only upper replacements.
  Preserve null/array replacement, accepted types, error boundaries, and input
  isolation. This directly removes the repeated descendant cloning measured in
  the assessment without an index or framework.
- **Namespace listings (20):** Resolve scope in the engine, then pass private
  `configuration.IdentityFilter{All bool, Namespace string}` to captured
  inventories. All copies the sorted slice once; named selection binary-searches
  the exact `namespace + "/"` prefix range and copies only matches. Unqualified
  selection scans immutable state directly, without first copying everything.
  Update the private inventory interface/fakes for backend/profile IDs and
  prompt identities. Preserve final context checks, ordering, exact prefix
  matching, absent results, view/default/All scope, and mutation isolation.

Existing sorted inventories supply named ranges without persistent buckets or
extra construction traversal. Buckets could reduce repeated lookup work but add
retained state even for consumers that never list resources. Unqualified entries
need not form a contiguous range; their selection remains a scan. No new public
signature, cache, pagination, or complexity guarantee accompanies this remedy.

The [baseline measurements](#planning-evidence) support implementing 17, 19, and
20 rather than deferring them. Repeat comparable cases after implementation;
expected savings and algorithmic estimates are not measured replacement results.
No exact peak-heap promise, benchmark threshold, or recurring release gate is
introduced. If new evidence changes the tradeoff, record it and the proposed
disposition rather than silently dropping selected work. The companion plan
coordinates the related traversal and trusted-copying changes in E.

### E. Internal Coordination And Test Assembly

**Catalog:** 22, 23, 24, 25, 34. **Decision:** 6.

**Current gap:** Large functions combine distinct responsibilities, JSON copying
is duplicated across boundaries with different policies, and a test-only legacy
capacity wrapper remains a production constructor.

**Work:**

- Extract focused configuration validation/materialization helpers with visible
  inputs and ownership. Preserve final-handoff validation, presence/null rules,
  origins, and meaningful error precedence without pinning incidental ordering
  of independent failures. Keep `validateEffective`'s resource/index traversal
  visible and extract backend/profile/prompt validators with record/path/index
  inputs. Keep typed presence interpretation together in profile construction.
  Split prompt capture into input/metadata projection, message content/template
  capture, and output-contract conversion; retain lazy origins, source-relative
  reads, and one final inspection/hash assembly.
- Separate requirement recognition/evaluation from reachability and evidence
  bookkeeping, sharing `if`/`with` orchestration where appropriate. Preserve
  scope differences, inactive computed-name behavior, and independent comparison
  with the actual renderer. Share the existing `parse.BranchNode` protocol,
  handling With's changed body dot and unchanged else dot explicitly. Preserve
  known/skipped/unknown classification and limitation bookkeeping; no generic
  visitor or additional template syntax is needed.
- Consolidate structural copying only where policies genuinely agree. Keep
  facade projection and runtime validation distinct; preserve existing types,
  nils, supported numbers, schema keys, and ownership. Infallible projection
  must not acquire new errors, panics, or silent fallbacks. Retain separate walkers
  if sharing would require a configurable reflection framework. B's removed
  source compatibility adapter must not migrate into this shared machinery.
  Put infallible cloning in `internal/jsonvalue.CloneTrusted(any) any` and route
  facade nested values through it. Runtime `Copy`/`CopyMap` validate, normalize
  indirection/null, and enforce limits/key policy; trusted cloning preserves
  accepted values and existing fast-path/reflection type/nil behavior without
  interpreting numbers. Do not substitute fallible `Copy` and hide errors.
  This centralizes ownership, not necessarily the traversal algorithm: two small
  walkers remain appropriate where their policies differ.
- Separate generation attempt results and accounting from dispatch/retry control
  flow with small internal helpers. Preserve built-in/injected differences,
  permit release before callbacks or backoff, frozen requests, existing retry
  classification, private evidence, and known/unknown usage distinctions.
  Replace execute's parallel returns with `attemptResult` containing
  `Outcome llm.AttemptOutcome`, `Dispatched`, `CapacityWait`, `RequestDuration`,
  and `StartedAt`. Share usage/evidence/event projections only where facts match;
  built-in retry and injected one-shot control flow remain distinct. C's
  preparation-failure marker is independent of attempt accounting.
- The initial/repair runner scheduling test uses the production generation
  coordinator via `generation.NewInjected` with the same manager. The production
  `internal/capacity/client.go` wrapper and its retention prose are retired.
  Preserve useful FIFO, cancellation, independent-pool, and panic-release scenarios with their
  owning components. Keep any useful one-shot scaffolding test-only; do not wrap
  the coordinator in another permit-acquisition layer.

**Completion evidence:** Existing meaningful behavioral tests and focused cases
for changed seams continue to exercise real internal collaborators. Tests no
longer require the retired production wrapper. Review demonstrates simpler
ownership/control flow; lowering complexity metrics alone is not an acceptance
criterion. No capacity-manager redesign or public DTO unification is needed.

### F. Documentation, Representation, And Test Maintenance

**Catalog:** 12, 13, 14, 28, 29, 30, 31, 32, 33.

**Current gap:** Several comments contradict implemented behavior, expectations
are duplicated or over-specified, and internal fields/helpers have no consumer.

**Work:**

- Correct credential-inspection GoDoc and its example: explicit profile credential
  environment selection can coexist with required credentials; only the
  backend-derived default is suppressed in that case. Correct provider-map wire
  documentation to distinguish configured recursive overlays from runtime
  replacement. Remove stale trimmed-ID and prompt-version comments.
- Give complete bundled-catalog expectations one independently specified test
  owner. Retain smaller tests for distinct loading, wire, and smoke behavior,
  using controlled fixtures where bundled content is irrelevant. Keep the full
  independent oracle in `internal/configuration/bundled_test.go` and replace its
  formatted-map plugin comparisons with structural assertions. Public tests own
  toggles/order/isolation/overrides; use a controlled model/plugin wire case and
  a configured-default smoke case instead of repeating the complete matrix.
- Replace literal artifact-hash algorithm assertions with documented opaque
  equality/ownership and byte/size behavior; keep benchmark setup outside timed
  work. Use same-body inline/file equality for size-hint cases and prepare an
  opaque expected hash outside benchmark timing. This does not require changing
  the artifact hash algorithm.
- Remove unused backend presence records and `QueueCapacitySet`, preserving
  effective queue defaults, explicit zero, provenance, and needed profile/input
  presence tracking. Remove the unused backend field list/map, resolution capture,
  and runtime queue-presence flag, not ConfigKit provenance.
- Remove duplicated captured prompt output-format and unused message-file fields.
  Preserve actual `content_file` loading, origins, and override-default behavior.
  Account for observable definition-hash changes without promising permanent
  encoding or bit-identical hashes across this change. Remove the unreachable
  fallback with the fields, preserving empty-format request defaulting and
  inspection/preparation equality.
- Remove unused private reachability endpoint state and the unused schema-helper
  result, preserving actual report evidence and public single-file behavior.
  Remove orphan `filecatalog.RelativePath`, `resourceNamespaceFromID`, and the
  unused runner-test `stringPtr` helper with their obsolete self-tests.

**Completion evidence:** Documentation agrees with the owning GoDoc/format/wire
contracts. Caller checks precede deletion; meaningful discovery, identity,
namespace, schema, privacy, and bundled behavior remain covered. Tests do not
merely reproduce the loader's output as their expected answer.

## Implementation Plan

The [companion plan](implementation.md) defines the ordered, prompt-sized
implementation units, dependencies, and validation for one end-to-end sprint.
Workstream letters group target outcomes, not delivery stages. Temporary
intermediate compilation gaps are acceptable; completion concerns the combined
end state. No significant product decisions remain open. Equivalent private
coding details remain flexible; new evidence requiring a scope or contract
change should be surfaced rather than silently changing the selected target.

## Compatibility And Documentation Delivery

The selected ownership, schema, and convenience-source changes are deliberate
pre-1.0 compatibility changes. Deliver them in a minor release under the existing
[release policy](../policy/release.md#versioning-and-compatibility), with migration
notes explaining independent artifact storage, self-contained schemas, and
ConfigKit-supported input types. Mention changed definition hashes so consumers
can account for comparisons or caches; do not establish a stable encoding.
Release selection and publication are not part of writing this roadmap.

Update exported GoDoc, [consumer guidance](../consumers/pkg-promptkit.md),
[formats](../formats.md), [wire integration](../integrations/openai-compatible-chat.md),
and relevant [internal documentation](../internal/overview.md) when their owning
behavior lands. Clarify ordinary development toolchain selection in
[development guidance](../development.md), linking to existing release policy.
Correct standalone documentation defects without waiting for unrelated code.

## Planning Evidence

These are historical assessment results from unchanged production code at the
baseline above, using Go 1.27.3 on linux/amd64, Intel i5-7260U. Assessment records
were completed in commit `3ef2a9e`. Dependency/source inspection supplied the
ownership, caller, and interface evidence; temporary offline probes supplied the
schema results and measurements below. No replacement implementation was built
or benchmarked. Aggregate allocation is not peak live heap, and noisy timings
are not throughput promises.

The temporary harness directory was `/tmp/promptkit-assessment-aqn3B8`. Its
continued existence is not a planning dependency: the reproduction recipes below
retain the workloads and invocation settings. Recreate targeted harnesses outside
the repository or carry equivalent cases into focused implementation benchmarks,
not a new permanent benchmark framework. Reuse unchanged evidence unless a
changed dependency, workload, or implementation makes a new measurement useful.

### Schema Dependency Probe

With jsonschema/v6 6.0.2, decode each document with `json.Decoder.UseNumber`,
create a fresh compiler with `DefaultDraft(Draft2020)`, register the document at
`https://example.invalid/retrieval`, and install a loader that counts requests
and returns an external-resource error. Compile the retrieval URI; for the
unused-definition case, additionally compile `#/$defs/unused` with the same
compiler. The harness used `go run -mod=readonly` with
`GOTOOLCHAIN=local GOWORK=off GOPROXY=off`; no network resources were fetched.

| Document case | Root compilation | External loads |
| --- | --- | --- |
| `$defs.v` string schema referenced by `#/$defs/v` | Succeeds | 0 |
| `$anchor: v` in a definition, referenced by `#v` | Succeeds | 0 |
| Object definition with `next` referring recursively to that definition | Succeeds | 0 |
| Explicit absolute root ID and reference to its definition through that ID | Succeeds | 0 |
| Embedded definition with relative `$id: child`, referenced as `child` | Succeeds | 0 |
| Root `$ref: child.json` with no embedded resource | Fails | 1 |
| Root string schema with unused `$defs.unused: {$ref: child.json}` | Succeeds; explicitly compiling the definition fails | 0, then 1 |
| Example instance object containing literal `$ref: child.json` | Succeeds | 0 |
| Standard 2020-12 `$schema` declaration | Succeeds | 0 |

Explicit-ID cases also worked when retrieval URI and document ID differed. A
review rerun reproduced these results. This supports B's rejecting loader plus
schema-location enumeration, not a lexical URI ban. Dynamic references, escaped
pointers, embedded dialect handling, and full location-table coverage remain
implementation regression cases, not claims established by this small probe.

### Overlay Baseline

Construct maps outside timing. Start with
`map[string]any{"leaf": 1, "array": []any{1, 2, "x"}}`; wrap it `depth` times in a map with
`child` pointing to the previous level and `width` scalar keys named `key0`,
`key1`, etc., each with its index value. Build lower and upper independently.
For the disjoint case, wrap upper once more under `new`; otherwise use matching
keys. Call unchanged `overlayProviderMap`, retaining its result and checking
errors. These are chains with scalar siblings, not exponentially branching trees.

Run a temporary test overlay in `internal/configuration`, substituting its
absolute JSON mapping path:

```sh
go test -mod=readonly -overlay=/absolute/path/overlay.json ./internal/configuration \
  -run '^$' -bench '^BenchmarkAssessmentOverlay$' \
  -cpu=1 -benchmem -benchtime=50x -count=3
```

The overlay maps an absolute package-local test filename to an external harness
file declaring that package; it does not edit repository source. The same
`GOTOOLCHAIN=local GOWORK=off GOPROXY=off` settings apply to all measurements.

| Depth / width / overlap | Observed ns/op range | B/op | Allocs/op |
| --- | --- | --- | --- |
| 2 / 2 / overlapping | 20,104–34,618 | 9,496 | 236 |
| 12 / 2 / overlapping | 263,387–287,814 | 117,869 | 2,609 |
| 40 / 2 / overlapping | 2,355,659–2,580,220 | about 1,241,620 | about 22,672 |
| 4 / 16 / overlapping | 168,483–178,889 | 65,434 | 1,591 |
| 12 / 2 / disjoint upper root | 49,437–53,966 | 32,496 | 686 |

Source inspection shows lower descendants cloned again at each recursive
overlap. The selected owned-overlay algorithm removes that repeated work;
post-change measurements must establish actual savings.

### Repair Feedback Baseline

Call unchanged `formatRepairDiagnostics` with one-element string slices:
64 ASCII `x` bytes, 1 MiB of `x`, and 1 MiB of repeated `<&>` plus NUL (four bytes
per repetition). Build inputs outside timing, retain the output in a sink, and
record input/output lengths after stopping timing. Use the overlay method above
in `internal/usecase`, benchmark `BenchmarkAssessmentFeedback`, with
`-cpu=1 -benchmem -benchtime=20x -count=3`.

| Input | Observed time per operation | Allocated bytes per operation | Output bytes |
| --- | --- | --- | --- |
| 64 ASCII bytes | 780–1,198 ns | 224 | 68 |
| 1 MiB ASCII | 5.83–7.99 ms | about 6.81 MB | 65,536 |
| 1 MiB escaped content | 27.7–39.1 ms | about 44.12 MB | 65,531 |

Full-string normalization and repeated full marshaling account for the observed
allocation scale. D's bounded-prefix design must preserve the existing fitting
result and omission semantics; these timings do not measure the new design.

### Listing Baseline

For N = 32, 1,024, and 4,096, create sorted IDs `namespace/p%06d` outside timing.
Sparse mode assigns one ID to `target`; dense assigns half. Other IDs use `other`.
All mode uses the sparse inventory with the All filter; absent mode selects
`missing`. The baseline list callback returns `append([]string(nil), ids...)`,
matching configuration's full-inventory copy. Invoke unchanged
`filterIdentityList` with background context and empty view namespace. Use an
`internal/engine` overlay, `BenchmarkAssessmentListing`, and
`-cpu=1 -benchmem -benchtime=100x -count=3`.

| Total IDs | Mode order | ns/op ranges in mode order | B/op | Allocs/op |
| --- | --- | --- | --- | --- |
| 32 | All / sparse / dense / absent | 1,176–1,332 / 759–937 / 1,053–1,875 / 797–819 | 1,472 / 528 / 960 / 512 | 5 / 2 / 4 / 1 |
| 1,024 | All / sparse / dense / absent | 41,851–48,965 / 13,860–19,334 / 38,416–45,031 / 15,258–19,282 | 78,144 / 18,448 / 37,184 / 18,432 | 11 / 2 / 9 / 1 |
| 4,096 | All / sparse / dense / absent | 194,479–202,439 / 92,265–95,215 / 148,209–151,180 / 93,157–94,526 | 305,472 / 65,552 / 166,208 / 65,536 | 14 / 2 / 12 / 1 |

This measures filtering/copying, not full construction or public projection.
For post-change comparisons, exercise D's selection-before-copy path rather than
retaining an artificial full-copy callback. Repeated requests repeat the
baseline full-inventory allocation. Binary search over existing named ranges
has estimated comparison work proportional to log N plus output and introduces
no persistent index; no alternative layout was benchmarked.

### Requirement Evidence Work Count

For a left-heavy recognized condition with N distinct missing names, the current
merge copies growing sets: `2 + 3 + ... + N` insertions, excluding leaf creation.
That is 9 at N=4, 135 at N=16, and 8,255 at N=128; the ordinary two-name case
needs two. These are source-derived counts, not elapsed-time measurements. The
selected small-to-large owned merge addresses that specific repeated copying;
it does not establish a universal inspection complexity contract.

## Validation And Completion

Follow the [testing policy](../policy/testing.md) and existing
[maintainer validation workflow](../development.md#maintainer-validation).
Use focused behavioral regressions for demonstrated defects, race testing for
changed concurrent ownership/scheduling paths, and offline wire tests where
local success does not prove the emitted request is correct. Reuse successful
checks for unchanged content; release validation remains the existing workflow.
No live inference, provider credentials, generalized benchmark suite, or new
recurring release gate is introduced.

The workstream catalog lists partition all 34 entries. During implementation,
record for each workstream its completed units, evidence, and any explicitly
justified deferral. A measured deferral is not an implemented fix; optional
cleanup does not hold unrelated fixes hostage. Full implementation completion
requires the selected changes, their relevant validation, migration guidance,
and removal of obsolete documentation/tests. Any remaining work must be named
and retained in an appropriate roadmap rather than lost during consolidation.

When this roadmap's work is complete or otherwise explicitly disposed of, move
still-useful contract information to its canonical owners and retire temporary
planning documents under the [documentation lifecycle](../policy/documentation.md#document-lifecycle).
Keep audit evidence and decision provenance discoverable while they remain
useful; do not maintain competing current-behavior references.