v2.6.0 - The lean core release

The most common feedback on the skill was that SKILL.md had grown too big, and the feedback was right: the whole file loads into the AI's context every time the skill triggers, and at 16,000 words it was starting every session deep in the hole. This release makes SKILL.md practice the progressive disclosure it preaches.

Nothing was deleted. Every gate, formula, script, and conversational move survives word for word. What moved was the reference-grade detail, into the reference files that already existed for exactly this purpose (much of it was duplicated there all along).

SKILL.md: 16,149 words / 96KB down to ~13,000 words / 78KB. Every session now starts thousands of tokens lighter, which means more room for the session itself and better attention on the rules that stay.

Added

• references/PLAN-TEMPLATE.md: the verbatim blocks for the Phase 8 plan document in one paste-ready file: the House Rules block, the checkpoint block format with its five rules, the three pre-launch audit prompts, the Working With Your AI Tool checklist, and the debug-logging prompt.
• The three loop shapes, walked out loud, in references/GROWTH-LOOPS.md: the how-it-spins narratives and worked call examples moved there from Phase 6.6, next to the taxonomy they belong with.

Changed

• Phase 0 points instead of repeating: the fetch ladder, the five sorting lenses, and the source-verification protocol were fully specced in references/DISCOVERY-DEEP-DIVE.md and re-taught in SKILL.md. The skill file keeps the beats, the gates, the scoring formula, the anchors, the evidence floor, and the no-go script.
• Phase 6.6 keeps the diagnostic, the references keep the teaching (GROWTH-LOOPS.md and COLD-START.md).
• Phase 8 keeps the outline, PLAN-TEMPLATE.md keeps the blocks.
• Repeated statements now appear once, with short callbacks.

Full details in the CHANGELOG.