Split SKILL.md (1,114 lines) into a core skill + on-demand references/ for progressive disclosure

[vickaul-ai]

The installed Claude Code skill (~/.claude/skills/graphify/SKILL.md, v0.8.14) is a single
1,114-line file. Claude Code loads the whole body every time the skill fires, and its trigger is
deliberately broad ("any question about a codebase … especially if graphify-out/ exists"), so a
quick one-line graph query pays the context cost of the entire pipeline documentation
(~10–12k tokens).

Agent Skills' progressive-disclosure convention handles this: keep the always-needed core in
SKILL.md and move branch-specific documentation into references/ files that the model reads
only when that branch is taken.

Suggested split (v0.8.14 line ranges):

Content	Lines	Destination
Usage, "what graphify is for", pipeline Steps 0–9	1–~765	keep in SKILL.md
Subagent extraction-prompt template	~332–397	references/extraction-prompt.md
--update incremental rebuild	~786–943	references/subcommands.md
--cluster-only	~947–955	references/subcommands.md
query / path / explain / add	~959–1053	references/subcommands.md
--watch	~1055–1074	references/subcommands.md
git hook integration	~1076–1088	references/subcommands.md
CLAUDE.md integration	~1092–1105	references/subcommands.md

Each moved section is replaced in SKILL.md by a one-line pointer, e.g.
For --update / --watch / query / git-hook / CLAUDE.md integration, read references/subcommands.md.

This keeps the common invocation (build or query a graph) at roughly 60% of the current size
while making the rarely-used branches free until needed. Happy to PR this if the structure is
agreeable.