Audit CONTEXT.md against the ADRs

Periodic maintenance. The ADRs (docs/adr/) are the source of truth; CONTEXT.md is the glossary that must track them. Goal: find and fix drift — stale terms, wrong citations, contradictions, implementation-detail leakage.

Workflow

1. Survey. Read CONTEXT.md; list docs/adr/*.md (plus docs/migration/, docs/agents/, docs/testing.md). Note the glossary's top-of-file meta-rule (no schemas / file paths / implementation details).
2. Group. Partition the ADRs into ~5 thematic groups (e.g. lesson & playback core; toy runtime contract & validation; registry / resolution / hosting / failures; identity / auth / multi-tenant / agent; tooling & conventions). Map each glossary term to a group.
3. Fan out. Launch one read-only explore sub-agent per group, in parallel (one message, multiple Task calls, readonly: true). Give each its ADR file list, its glossary terms, and the per-group checks below. Subagents do NOT have the chat context — spell out every file path and term.
4. Synthesize. Merge findings into ONE prioritized drift report: (1) contradicts an ADR decision, (2) wrong/stale ADR citation, (3) glossary-rule violation (impl-detail leakage), (4) missing canonical term, (5) ADR-vs-ADR conflict. Link sub-agents by id. Present to the user before editing.
5. Verify before editing. For each finding, open the cited ADR and confirm against its actual text — never trust a sub-agent summary alone (per AGENTS.md: verify before recommending). Note what you couldn't confirm.
6. Apply in-scope fixes only. Factual corrections, citation fixes, and stale-term renames: apply. Scope changes (stripping implementation detail, adding new terms): ask first. ADR-vs-ADR conflicts: do not silently pick a winner in the glossary — file an issue or draft an ADR amendment, then make the glossary follow that decision.

Per-group checks (paste into each sub-agent prompt)