Implementation

Progress

[###-] Phase 4/5 - deterministic fixture and scratch-composer evidence complete; manifest decisions and durable promotion remain open.

Implemented

Added nine durable Harborworks fixture repositories under the dated investigation, with 54 shallow skills and mixed dependency-free Python and Node CLIs.

Added order, revision-conflict, policy-conflict, and two-function/three-domain prefilter scenarios.

Implemented the ignored scratch composer under .cache/src/harborworks-composition/.

Generated full, repository-filtered, artifact-filtered, and naive-overlay raw results under .cache/results/harborworks-composition/run-001/.

Recorded the reviewed observations and interpretation boundary in the dated investigation.

Recorded 15 evidence-linked candidate rules in the investigation without creating the deferred durable specification.

Added a full-union materializer that writes the harness surface directly to .cache/harborworks-operations-workbench/{skills,knowledge_base,src}.

Prefixed all generated skill folders and frontmatter names with their source repository, and namespaced knowledge and tooling one level below their common roots.

Rebased the fixture's known knowledge links, CLI entrypoints, skill names, knowledge arguments, and execution-root wording while preserving source and materialized digests in the generated catalog.

Validated 54 skill discoveries, 54 relative links, and 54 materialized CLI invocations, then recorded the limitations in Run 002.

Deliberate deviations

No durable specification was created before experimentation because the maintainer explicitly deferred it until the manifest semantics are clearer.

No implementation was promoted into committed src/; the experiment stays on the .cache/src side of the code-promotion ladder.

Fixture revisions are labels rather than nested Git repositories so the investigation can preserve inputs without committing nested .git state.

Remaining decisions

Artifact identity and the normative role of content digests.

Export include/exclude rules and artifact classes.

Manifest versus lockfile responsibility for immutable resolution.

Optional repository scope hints versus required prefilter semantics.

Effective policy calculation across paths and artifact classes.

Whether a model/harness benchmark should precede the first durable spec.

Portable reference restrictions versus generated compatibility wrappers.

Whether .composed/ should become the standard generated surface.