v2.6.1

Patch release: platform parity and internal consistency. No new scripts, skills, or features.

Systemic Honesty is no longer Claude-only

The three core principles — never fabricate, skill-first, methodological rigor — were injected only into Claude sessions, via its SessionStart hook. Codex, Gemini, and OpenCode loaded the skills but not the constraints that govern them. All four platforms now carry the same principles, byte-identical, plus the rule that no bibliography ships unverified and the uv prerequisite for the research toolchain.

Fixes

• grant-writing name mismatch — the skill's frontmatter said grant-proposal while its directory and every doc said grant-writing. Claude keys by directory so it was unaffected, but Gemini keys by frontmatter name and registered it wrong, so invoking grant-writing on Gemini failed. Found by actually running gemini skills list to verify a documentation claim.
• Deleted .codex/skills/ — eight dead duplicate skill files the Codex launcher never loaded (it reads skills/ directly) and which had drifted badly stale.
• Rewrote GEMINI.md — it had a duplicated section, listed 9 of 15 skills, and pointed at an agents/ directory that does not exist.
• Removed dead code from the OpenCode plugin.
• Corrected KnowledgeGraph.md (claimed three CLI platforms, omitted OpenCode) and the README architecture section.

125 tests across 9 files.