fix: remove the consumer name that shipped in vendored bytecode

[baderdean]

Follow-up to #128, which left .jcode/ out of the consumer-name cleanup. That exclusion was wrong, and the reason I gave for it was wrong twice over.

The reasoning I got wrong

I claimed .jcode/skills/ is "vendored upstream and re-synced by bin/update, so an edit there is churn the next sync reverts". bin/update's own comment settles it:

.jcode/ is synced as a whole — the engine owns it entirely (agents/, skills/, UPSTREAM-FIX-WORKFLOW.md, DESIGN-template.md, prompt-overlay.md)

The sync runs from this repo to consumers. This repo is the origin, not a vendored copy: an edit here is the fix, and it propagates on the next engine bump. The Makefile does exclude .jcode/ from shfmt, but that is about reformatting churn in vendored scripts — not a reason to leave content unreviewed.

What was actually in there

A consumer's project name, in an example command. .jcode/skills/ui-ux-pro-max/SKILL.md shipped:

python3 …/search.py "humanitarian NGO editorial brutalist dark" --design-system -p "<consumer project>"

Every agent that loads that skill reads it. It was the most explicit consumer name left in the repo, and #128 walked past it.

The same name in tracked bytecode. Four __pycache__ files were committed, and two carry that project name in bytecode compiled from a source version that no longer contains it:

.jcode/skills/ui-ux-pro-max/scripts/__pycache__/core.cpython-39.pyc
.jcode/skills/ui-ux-pro-max/scripts/__pycache__/core.cpython-314.pyc
.jcode/skills/ui-ux-pro-max/scripts/__pycache__/design_system.cpython-39.pyc
.jcode/skills/ui-ux-pro-max/scripts/__pycache__/design_system.cpython-314.pyc

A leak no text grep finds, shipped to every consumer by bin/update (SYNC_PATHS includes .jcode). The skill's .gitignore already listed __pycache__/ and *.pyc — a tracked file ignores .gitignore, which is how they survived it.

Byte-compiled output is regenerable, unreviewable, and outlives the source it was built from. Untracked and deleted.

The guard

A fourth test: no tracked compiled artifact. That was the vector, and unlike the name itself it is mechanically checkable.

Scope is now split rather than uniform:

check	scope	why
host allowlist	engine surface, minus .jcode/skills/	those skills cite framework docs, vendor design systems, security references — some forty domains. A host there is a citation, not infrastructure; an allowlist would be forty entries long and would flag every reference a skill adds.
cross-project reference	everything, .jcode/ included	does not fire on citations
compiled artifact	everything	ditto

Two corrections to the guard, both caught by running it

• <owner>/<project>#N is no longer matched. It cannot be told apart from citing a real upstream bug in a dependency, which is provenance a comment should carry — docker/Dockerfile.agents cites one in its base image, and widening the scope surfaced it as a false positive. Every consumer leak actually found was host-qualified.
• The file scans itself, so the literal hostnames I wrote as examples in its own comments failed its own host check. Second time this has bitten (fix(dispatch,ci,diagrams): opt-in issue entry, stop polluting consumer CI on both forges, legible diagrams, no consumer names in the engine #128 hit it in CI with something.dev). There is now a note in the file: name domains by description, not by spelling them.

What it cannot catch, stated in the file rather than glossed

An arbitrary proper noun. No pattern separates a consumer's project name from any other capitalised phrase without the blocklist this design exists to avoid — a blocklist would have to spell out the identities the cleanup removes. What is checkable is the vector it travelled by.

Tests

make check green: 1117 tests, shellcheck + shfmt + check-lessons clean, bin/check-doc-sync clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HsXCEtWaFU9nWvKziiNqqy

---

Generated by Claude Code