Correct Governed Memory (Taheri) attribution; print arXiv IDs in refs

[mthamil107]

What

Removes the fabricated "Co-memorize" pattern attribution to Taheri's Governed Memory (arXiv:2603.17787). That paper describes automated governance for autonomous agents with no human-in-the-loop gate, so crediting it with memorywire's diff-and-approve HITL pattern was a misattribution. Reframed as a contrast everywhere — which also makes memorywire's human-gated governance more clearly novel.

Changes

• Paper (memorywire-paper.md, .tex, arXiv bundle .tex): reframe all occurrences (§1.1, §1.4, C3, §2.5, §4.5, Acknowledgments) as contrast, not precedent.
• Bibliography (memorywire.bib ×2): plain.bst silently drops eprint/archivePrefix/primaryClass, so no arXiv ID was printing in the reference list. Added note = {arXiv:ID} to all five arXiv entries so every reference is followable. Fixed Taheri primaryClass cs.MA → cs.AI; removed the stale % TODO verify comment.
• CITATION.cff: replaced the invented title and placeholder author "Governed Memory authors" with the real values (Hamed Taheri; "Governed Memory: A Production Architecture for Multi-Agent Workflows").
• README + launch assets (docs/launch-post.md, docs/launch/01-tweet-thread.md) + reference bullets (docs/MCP-RELATIONSHIP.md, docs/spec/v0.md): same contrast reframe, so no publishable copy of the error remains.
• Regenerated docs/paper/arxiv-submission.tar.gz.

Notes

• The arXiv v4 replacement carrying these corrections is already submitted.
• co-memorize still appears as the governance UI's own screen/route name (/co-memorize) — that's product naming, not a citation, and is intentionally left unchanged.
• Docs-only change; no code touched, so lint/type/test CI is unaffected.