Use host-generic graphify skill instructions

[ari-mitophane]

Purpose

Fix Graphify-generated install guidance so Codex/agent-facing surfaces do not tell hosts to invoke a literal skill tool with skill: "graphify". The generated wording now stays host-generic and points users to the installed graphify skill or instructions.

Implemented Delta

• Updated the always-on AGENTS source fragment, packaged graphify/always_on/agents-md.md, and expected skillgen snapshot to use host-generic /graphify guidance.
• Updated _skill_registration() so Claude/CodeBuddy registration snippets use the same host-generic wording.
• Replaced stale English code-block examples in translated README files.
• Added regression tests for AGENTS and registration helper wording.
• Tightened the always-on roundtrip test to allow only the exact approved agents-md sentence delta.
• Made skillgen git blob reads decode as UTF-8 explicitly on Windows.

Current Behavior

graphify codex install and related generated/registration examples can surface wording that tells agents to invoke a literal skill tool with skill: "graphify", which is host-specific and not valid in every environment.

Desired Behavior

Generated install guidance should say that /graphify should use the installed graphify skill or instructions before doing anything else, without naming a literal tool API that may not exist.

Key Interfaces / Modules

• tools/skillgen/fragments/always-on/agents-md.md: source AGENTS wording.
• graphify/always_on/agents-md.md: packaged always-on artifact.
• tools/skillgen/expected/graphify__always_on__agents-md.md: generated snapshot.
• graphify/__main__.py: _skill_registration() output and UTF-8 git blob reads.
• tests/test_install_strings.py and tests/test_skillgen.py: regression coverage.

Decision Path

The bad text is produced by Graphify install surfaces, so the durable fix is upstream in the template/generator path rather than downstream edits in every consuming repository. The historical roundtrip guard remains strict: the test now proves the rendered AGENTS block equals the v8 baseline with only the exact approved sentence replacement.

Discarded Candidates

• Downstream AGENTS edits were rejected because graphify codex install would regenerate the old wording.
• A file-level agents-md roundtrip exemption was rejected because it could hide unrelated drift.
• Broad translation rewrites were rejected because the stale translated README lines are English code-block examples.

Verification

• python -m pytest tests/test_skillgen.py::test_always_on_roundtrip_is_byte_faithful -q
• python -m pytest tests/test_install_strings.py -k "agents_section or skill_registration"
• python -m pytest tests/test_install.py -k agents_install
• python -m pytest tests/test_skillgen.py -k always_on
• python -m tools.skillgen --check
• rg -n 'invoke the Skill tool|invoke the skill tool|skill: "graphify"|skill tool' -S . --glob '!tests/**' --glob '!graphify-out/**' --glob '!.git/**'
• git diff --check
• graphify update .

Noted but not caused by this PR: broader local pytest over tests/test_install_strings.py tests/test_install.py tests/test_skillgen.py has pre-existing Windows/source-checkout failures around pytest import/source state, Windows path separator expectations, and missing installed graphifyy package metadata causing __version__ to be unknown.

Failure Modes / Residual Risk

Line-ending warnings appear in this Windows checkout because global core.autocrlf is enabled. The committed diff is limited to the intended wording, test, and generator changes.

Reviewer Guidance

Start with tests/test_skillgen.py::test_always_on_roundtrip_is_byte_faithful to check the intentionally narrow historical-baseline exception, then review the AGENTS fragment and _skill_registration() wording.

Out of Scope

• Changing hook behavior.
• Reworking translated README prose beyond the stale English code-block sentence.
• Fixing unrelated Windows/source-checkout test failures.

Related Issues

None.

[safishamsi]

Merged into v8 as e3e4198 with your authorship. The host-generic wording is the right fix — telling agents to invoke a literal skill tool with skill: "graphify" only made sense on Claude and was invalid elsewhere.

One thing I had to add before it could land: the change passed skillgen --check and the pytest mirror, but the authoritative --always-on-roundtrip CLI guard (a blocking CI step) was still red, because the PR only patched the pytest test to tolerate the mismatch and not the guard itself — so the two disagreed. Rather than weaken the guard, I added an explicit, reviewable ALWAYS_ON_SANCTIONED_EDITS registry: the guard now applies the approved old→new substitution to the frozen baseline before the byte compare, so this exact sentence is allowed while any other drift still fails. CLI guard and pytest test now agree and all five skillgen guards are green. Thanks for the clean, well-scoped PR.