Track post-rollout doctor warning cleanup for fleet PGLite installs

[100yenadmin]

TL;DR

The eva-v0.41.18.5 fleet rollout passed runtime gates on every active VM, but doctor scores remain below 100 because doctor includes PGLite applicability, graph/timeline/takes/content-density, and legacy embedding config warnings. This is not a release blocker, but we should clean up the warning model so fleet health is easier to interpret.

Evidence From Rollout

Runtime gates were green across Golden, Eric, Jackie, David, Matt, Tarzan, and evaos-support:

• gbrain version = 0.41.18.5
• DB opened
• openclaw-support-kb was populated and searchable
• openclaw-gateway active
• OpenClaw GBrain plugin loaded with expected tools and HTTP route
• duplicate stale gbrain serve process count = 0

Doctor scores still ranged from 55 to 75 on VMs because of non-runtime warnings:

• pgvector / jsonb_integrity on PGLite installs
• graph/entity/timeline/conversation coverage on sparse customer content
• takes_count when takes bootstrap is disabled
• content_sanity_audit_recent warnings in Support KB with hard=0 soft=0
• legacy embedding_env_override where the DB config row differs from the env-backed provider profile while runtime embedding still works

Acceptance Criteria

• Doctor output distinguishes runtime/install blockers from content-quality warnings.
• PGLite-only applicability checks do not make healthy fleet installs look broken.
• Existing homes with legacy embedding_model rows get a non-destructive remediation or a clearer warning.
• Updater/postcheck docs explain when to source /root/.gbrain/gbrain.env for direct embed repair commands.