Back up graph.html on both rebuild paths (CLI and skill)

[GUGU75006]

What

Two commits, one bug: graph.html was missing from the pre-overwrite snapshot that backup_if_protected() takes before a protected graph gets rebuilt.

1. graphify/export.py — add graph.html to _BACKUP_ARTIFACTS. This covers the CLI path (graphify update, watch._rebuild_code), which already calls backup_if_protected.

2. tools/skillgen/fragments/core/core.md — call backup_if_protected(Path('graphify-out')) in Step 4 of the generated runbook, right before to_json. This covers the AI-assistant skill path (/graphify --update for docs/papers/images), which builds the graph inline and never called backup_if_protected at all — not just missing graph.html, missing the whole snapshot.

Why

backup_if_protected() exists so a protected graph (semantic marker present, or curated community labels) survives a rebuild. Before this PR:

• CLI path: preserved graph.json, GRAPH_REPORT.md, manifest.json, cost.json, labels, analysis — but destroyed the interactive viewer on every run, and the dated graphify-out/<date>/ folders couldn't be opened in a browser at all.
• Skill path: preserved nothing. The generated runbook writes graph.json / GRAPH_REPORT.md / graph.html directly via inline Python and never invoked the backup helper, so a semantic (/graphify --update) run silently overwrote the entire prior snapshot — including the LLM-derived output that's expensive to reproduce, which is the exact case the helper was built to protect.

Both were found and reproduced on a real graphify-out/ (curated labels, no semantic marker) while investigating why a dated backup folder had no graph.html and the CLI update ran clean while the skill update visibly lost data.

Scope notes

• core.md is the shared fragment for every "split"-bucket platform (claude, codex, windows, opencode, kilo, copilot, claw, droid, amp, trae, kiro, pi, vscode) — 13 rendered skill files + their expected/ snapshots, all regenerated from this one source change.
• Left aider.md/devin.md untouched. Per the CHANGELOG, those are hand-maintained monoliths frozen against a pinned v8 blob by a round-trip guard with an explicit sanctioned-change allowlist; this change isn't in that allowlist and monkeying with it is out of scope here.
• Placement in Step 4: after the empty-graph guard (nothing worth protecting if extraction produced no nodes and we're about to abort) and before to_json (so the snapshot captures the previous run's graph.html, which Step 6 rewrites later in the same runbook).

Testing

• tests/test_export.py::test_backup_includes_graph_html (new) — asserts graph.html round-trips through backup_if_protected.
• Manual check against the installed skill fragment (isolated the Step 4 code block, substituted the documented placeholders, ran it against a synthetic graphify-out/ with curated labels):

  [graphify] backed up curated graph (4 files) -> 2026-08-06/
  

  graph.html, graph.json, GRAPH_REPORT.md all present in the dated folder afterward.
• Manual check against the CLI (graphify update . on a real project):

  [graphify] backed up curated graph (6 files) -> 2026-08-06/
  [graphify watch] Rebuilt: 3409 nodes, 10537 edges, 180 communities
  

• python -m tools.skillgen --check — OK, 125 artifacts match.
• python -m tools.skillgen --monolith-roundtrip — OK.
• python -m tools.skillgen --schema-singleton — OK.
• python -m tools.skillgen --audit-coverage — OK.
• python -m tools.skillgen --always-on-roundtrip — OK.

I could not run the Python test suite itself in this environment (pytest isn't installed against the interpreter the editable install uses) — please run CI on it.

🤖 Generated with Claude Code

[graphify-labs]

Graphify reviewed this change.

Looks safe to merge — no coupling regressions and no blocking issues, checked against the code graph (not a self-assessment).

---

Graphify review — findings

This pull request removes several repository files and trims documentation/CI configuration. It deletes the publish.yml PyPI publishing workflow, the BENCHMARKS.md file, and the .DS_Store entry from .gitignore, edits a comment in the CI security-scan job, and truncates a large portion of the CHANGELOG.md. The changed-symbols list also indicates edits across many test fixtures, extraction modules, skill-generation templates, and translated docs. Note: the diff was truncated in the prompt, so the full surface area (particularly the changes reflected in the symbol list) is not entirely visible here.

No blocking issues surfaced.

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 6277 functions depend on the 5912 functions this change touches.

Health — this change adds coupling hotspots:

• worse: main() — 55 callers, 109 callees
• new: extract_bash() — 19 callers, 8 callees
• new: to_html() — 14 callers, 9 callees
• new: install() — 12 callers, 10 callees
• new: _extract_generic() — 17 callers, 7 callees
• new: _resolve_js_module_path() — 28 callers, 4 callees
• new: codebuddy_install() — 21 callers, 5 callees
• new: extract_pascal() — 14 callers, 6 callees
• …and 49 more

Verification — 6277 functions in the blast radius were not formally verified this run (proofs are advisory here).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 6271 function(s) in the blast radius were not formally verified this run

· 57 more finding(s) on lines outside this diff (see the check run).

[graphify-labs]

Graphify reviewed this change.

Looks safe to merge — no coupling regressions and no blocking issues, checked against the code graph (not a self-assessment).

---

Graphify review — findings

This pull request removes several repository files, including the PyPI publishing GitHub Actions workflow (.github/workflows/publish.yml), the BENCHMARKS.md document, and the .DS_Store/macOS Finder entry from .gitignore. It also trims the top CHANGELOG.md entries (removing the unreleased 0.9.35 section and other content) and adjusts a comment in the CI workflow's security-scan job to mention the safety tool. The changed-symbols list suggests wider edits across source and test files, though the truncated diff only shows the workflow, gitignore, benchmarks, and changelog changes.

No blocking issues surfaced. 1 lower-confidence candidate did not survive cross-model review.

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 6277 functions depend on the 5912 functions this change touches.

Health — this change adds coupling hotspots:

• worse: main() — 55 callers, 109 callees
• new: extract_bash() — 19 callers, 8 callees
• new: to_html() — 14 callers, 9 callees
• new: install() — 12 callers, 10 callees
• new: _extract_generic() — 17 callers, 7 callees
• new: _resolve_js_module_path() — 28 callers, 4 callees
• new: codebuddy_install() — 21 callers, 5 callees
• new: extract_pascal() — 14 callers, 6 callees
• …and 49 more

Verification — 6277 functions in the blast radius were not formally verified this run (proofs are advisory here).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 6271 function(s) in the blast radius were not formally verified this run

· 57 more finding(s) on lines outside this diff (see the check run).

[graphify-labs]

Graphify reviewed this change.

Looks safe to merge — no coupling regressions and no blocking issues, checked against the code graph (not a self-assessment).

---

Graphify review — findings

This PR adds graph.html to the list of artifacts (_BACKUP_ARTIFACTS) that backup_if_protected snapshots into the dated graphify-out/YYYY-MM-DD/ folder before a rebuild overwrites a protected graph. It also inserts a backup_if_protected(Path('graphify-out')) call (and the corresponding import) into the export/build step across the various agent skill markdown files (amp, claw, codex, copilot, droid, kilo, and others) so the skill code paths perform the same snapshotting the CLI does. Accompanying changes update the CHANGELOG and expected skill-generation fixtures/tests to match. The surface area spans the core export.py, many skill-*.md runbook files, generated test fixtures under tools/skillgen/expected/, and a couple of export tests.

No blocking issues surfaced. 3 lower-confidence candidates did not survive cross-model review.

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 1353 functions depend on the 1006 functions this change touches.

Health — grade B; 10 existing hotspot(s) in the area this change touches (pre-existing, not introduced here):

• _rebuild_code() — 93 callers, 51 callees (high)
• build_merge() — 41 callers, 14 callees (high)
• to_obsidian() — 24 callers, 11 callees (high)
• to_json() — 42 callers, 6 callees (high)
• dispatch_command() — 2 callers, 117 callees (high)
• _make_graph() — 30 callers, 6 callees (high)
• run_pipeline() — 8 callers, 13 callees (high)
• to_canvas() — 14 callers, 3 callees (high)
• …and 2 more

Verification — 1353 functions in the blast radius were not formally verified this run (proofs are advisory here).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 1213 function(s) in the blast radius were not formally verified this run