docs: reframe org README around shared persistent workspaces

[chazmaniandinkle]

Summary

Replaces the old "AI agents need persistent memory" framing with a sharper thesis: every modern AI tool already has memory; what's missing is shared memory across tools, machines, and orgs. CogOS is the substrate underneath that solves that. Three substantive changes:

1. New opener. Leads with the cross-tool / cross-device / cross-org gap and the "hierarchical like git remotes, but recursive" framing. The persona is a Claude Code or Cursor user feeling fragmentation friction, not someone who needs to be told that AI tools forget things (most of them don't anymore).

2. New three-piece architecture diagram. AI tools → kernel → workspace overlay. Replaces the seven-module diagram, which referenced archived openclaw reconciler providers and tried to surface every implementation detail at once. New ".cog/ and .git/" section explicitly positions the cognitive overlay as a sibling to git, with content-addressed cross-references rather than duplication.

3. New 'Workspaces compose' section. Names the recursive upstream pattern as the operational expression of selective sovereignty: workspaces have upstreams the way git repos do; those upstreams have their own upstreams; promotion across the hierarchy is explicit. This is the load-bearing differentiator and was missing entirely from the old README.

What's cut

• TRM name and "Tiny Recursive Model" expansion (retired internal name)
• NDCG@10 numbers, LOC counts, test counts (volume metrics)
• Em-dash substitutes (`--`) in body prose, replaced with commas/periods/parens/colons
• Old seven-module diagram with archived openclaw references
• Em-dash in original body line ("any language — the wire protocol is the contract")

What's softened

• Constellation row in the repos table reframed as L1 trust-node protocol (aligned with cogos-dev/constellation#7, merged earlier today)
• Research row reframed away from EA/EFM/LoRO acronym soup, toward "notes and the training pipeline behind the kernel's design choices"
• Kernel features rewritten as a lean seven-bullet summary that names externalized attention as the conceptual frame and reconcilers as a primary kernel function

Alignment

This is the third repo in the docs-reframe sequence, all from 2026-04-29:

1. cogos-dev/constellation#7 — constellation README + PAPER (merged)
2. cogos-dev/cogos#84 — cogos README (merged)
3. This PR — org README

Net result: the three public-facing surfaces describe the substrate consistently. Visitors can land on any of them and not get a contradictory mental model.

Test plan

• Read top-to-bottom, confirm flow: pain → solution → diagram → overlay → hierarchy → kernel features → repos → quickstart
• Confirm no `—` or `--` in body prose
• Confirm no TRM, NDCG numbers, LOC counts, test counts
• Confirm constellation row matches the framing in cogos-dev/constellation
• Confirm `.cog/` and `.git/` sibling framing matches the cogos README