scripts/kg/ exists in template but is missing from the manifest; CLAUDE.md hints at it

[psaboia]

Defect

scripts/kg/ exists in the template repo with substantive content (build scripts, SPARQL queries, README, fixtures):

```
scripts/kg/build-graph.sh
scripts/kg/build-graph.py
scripts/kg/wiki-to-jsonld.py
scripts/kg/README.md
scripts/kg/sparql/.rq (13 sparql query files)
scripts/kg/fixtures/mini-wiki/ (test fixtures)
scripts/kg/build/ (runtime build artifacts, ignored)
scripts/kg/.cache/ (runtime cache, ignored)
```

But `scripts/kg/` is not enumerated in any array of `scripts/lib/template-manifest.sh` (introduced in PR #60), so neither `adopt.sh` nor `update-from-template.sh` nor `check-template-version.sh` sync those files. Meanwhile, `wiki/agents/claude-code/templates/claude-md-snippet.md` (which the overlay's `setup.sh` injects into derived projects' `CLAUDE.md`) references a `### Knowledge Graph` subsection, hinting at a KG/SPARQL pipeline that derived projects never receive.

How it surfaced

Manual end-to-end test during PR #60 review-resolution: adopted on a FUNSD scratch clone, opened Claude Code, ran `/wiki-experiment`. The agent honestly flagged the gap:

"CLAUDE.md references a `scripts/kg/` KG/SPARQL pipeline, but `scripts/kg/` does not exist in this adoption — KG rebuild not run, not verified."

The agent applied observe-the-failure discipline correctly. The defect is pre-existing (predates PR #60); the manifest just made the discrepancy auditable.

Design decision needed

Option A — ship the executable subset. Add to `TEMPLATE_SHARED_INFRA`:

• `scripts/kg/build-graph.sh`
• `scripts/kg/build-graph.py`
• `scripts/kg/wiki-to-jsonld.py`
• `scripts/kg/README.md`
• All `scripts/kg/sparql/*.rq` (13 files)

Plus gitignore `scripts/kg/build/` and `scripts/kg/.cache/` at the template level (or check that the existing `.gitignore` handles them; they may already be ignored). `scripts/kg/fixtures/mini-wiki/` is a test fixture — decide whether it belongs in the shipped set or stays template-internal.

After Option A, derived projects gain the KG pipeline that `claude-md-snippet.md` already hints at. The expectation in `CLAUDE.md` and reality on disk match.

Option B — remove the KG reference from the snippet. Edit `wiki/agents/claude-code/templates/claude-md-snippet.md` to drop the `### Knowledge Graph` subsection. `CLAUDE.md` in derived projects no longer makes a promise the template does not deliver. Smallest scope, smallest blast radius.

Option C — make KG opt-in via the features mechanism (RFC #13). Move `scripts/kg/` into `features/knowledge-graph/` with a `feature.json`, so adopters that want it run `./scripts/enable-feature.sh knowledge-graph`. Larger surface; aligns with the feature-flag direction.

Recommendation: Option B short-term (immediate consistency); Option C medium-term if the team wants KG to be a first-class opt-in capability. Option A is tempting but ships ~20 files most adopters do not actively want.

Acceptance

• After whichever option is chosen: a fresh adopt + Claude Code session on a non-template repo does NOT surface the gap (no honest "references X but X does not exist" flag).
• `scripts/test/tests/unit/manifest-shape/` continues to pass.
• If Option A: a new convergence assertion that `scripts/kg/` files reach the adopted host (the existing convergence smoke would catch this for free if the paths land in `TEMPLATE_SHARED_INFRA`).
• If Option B: any wiki page that links to the Knowledge Graph subsection of `CLAUDE.md` is updated to stop expecting it.

Context

Surfaced during PR #60 manual adoption test (Bloco 1 recipe documented in the PR body). The defect demonstrates that the manifest works as an auditing tool: discrepancies between "what `CLAUDE.md` tells users about" and "what the template actually ships" become findable. This issue is the first such finding.

Refs: PR #60 (manifest introduction; "Out of scope" section flagged this); `wiki/agents/claude-code/templates/claude-md-snippet.md`; RFC #13 (features mechanism, if Option C).

[psaboia]

New production evidence, from a /wiki-lint pass on PRISM (a repo without scripts/kg/):

1. Running wiki/init-wiki.sh in update mode injected the Knowledge Graph section into the main repo's CLAUDE.md — pointing at a scripts/kg/ pipeline the repo does not have (instantiate strips that section at bootstrap for exactly this reason, but update mode re-adds it unconditionally). The uncommitted hunk was discarded on review.
2. The /wiki-lint skill's step 7 ("Optionally rebuild the knowledge graph: ./scripts/kg/build-graph.sh") references the same non-existent script; the lint agent had to skip it and flagged it upstream.

So the gap now bites through two shipped surfaces (CLAUDE.md injection + lint procedure), not just the snippet this issue originally described. Worth factoring into the A/B/C decision: whichever option is chosen, update-mode's KG-section injection and the lint skill's step 7 need to become conditional on scripts/kg/ actually existing (or on the feature being enabled, under option C).

[psaboia]

PRISM adopted scripts/kg/ manually (crcresearch/PRISM PR #3, verified against its real wiki: 536 triples / 29 pages, SHACL clean) — second live consumer besides rare-disease-copy. +1 for option C (opt-in feature), which would let both converge through the normal sync path.

[psaboia]

Two packaging findings from PRISM's first real build-graph.sh run (PRISM PR #3), both relevant to whichever option (A/B/C) this issue lands on:

1. Remote shapes spec bug — now tracked upstream as LA3D/llm-wiki-colab#1 (sh:prefix "llm-wiki" vs llm-wiki-colab: constraints; pyshacl aborts). Template-side angle that remains regardless of that fix: the pipeline's only workaround is patching scripts/kg/.cache/shapes.ttl, which --refresh-spec silently discards. build-graph should degrade gracefully on a broken remote spec (skip validation with a loud note, or pin a known-good shapes copy) instead of aborting.
2. build-graph.sh assumes system python3. In uv-managed projects the deps live in the project environment, so the documented wrapper fails while uv run python scripts/kg/build-graph.py works. The wrapper should detect/prefer the project environment (uv/venv) before falling back to system python.

Both surfaced only on a real downstream repo — the template's own CI has the deps globally and its cache pre-populated.

[psaboia]

Third live consumer: AI-Sci-Disc-Mem has carried scripts/kg/ since its own June template sync (commit 421890a there, alongside wiki-write-protocol) and its wiki already ships an Edge-Types page. With PRISM (PR #3) and rare-disease-copy, that's three downstream repos maintaining manual copies of an unmanifested directory — the strongest signal yet for option C (opt-in feature) so all three converge through the normal sync path.