Harden read-only Web memory access and add per-record audits

[Conner512]

Summary

• Separate ChatGPT Web memory reads from local writes, project restoration and handoff.
• Enforce a default-deny, current-revision visibility boundary with an explicit local grant inventory and immediate request-time revocation checks.
• Add revision-pinned Unicode body/source pagination and bounded summary reads without exposing ineligible references.
• Add content-free, allowlisted request audits and an offline viewer, including conflict-only returned references.
• Document real-client acceptance, operational limits and public/private evidence boundaries.

Verification

• Supported Node 24 local matrix: 422/422 passed, with no failures, skips or cancellations.
• Deployment-host relevant subset: 80/80 passed; rollback backup retained, no database migration or unrelated configuration changes.
• Real ChatGPT Web: unauthorized search omission, six-page complete synthetic body read, fresh-conversation read, and per-memory grant revocation verified. Final deployed source also produced expected success/denial audit records.
• Source, index and selected publishing history checked with publication rules and pinned, redacted Gitleaks. Only synthetic fixtures and sanitized evidence are included.
• Corrected a publication-policy false positive for GitHub's exact public no-reply merge-commit email; regression coverage still rejects personal and lookalike addresses. No checks are bypassed.

Boundaries

No Web writes, project restoration, handoff, Organizer submission or new OAuth scopes. No actual model-provider or vector-service validation is claimed. OAuth connection identity is not physical-device identity; completed HTTP responses do not alone prove client consumption. Global production_ready remains false. Private runtime data, credentials, deployment files and evidence stay outside Git.