Not a demo.
The actual system.

What we knew before this run: From cmd-007 we knew our hooks don't return JSON stdout per spec — we identified the missing additionalContext fields. That was one finding from one observation.

What the fetch returned: Complete operational surface. 22 hook event types documented. We are using 3. That means 19 event types are completely unknown to our implementation. This is not a partial match — this far-exceeded expectations.

Key deltas from what we believed:
• Exit code behavior is event-specific — exit 2 on PostToolUse does NOT block (tool already ran). We had the wrong mental model.
• CLAUDE_ENV_FILE env var unknown to us — hooks can persist env vars for subsequent Bash commands.
• once: true field in skill frontmatter — hook runs once per session then removes itself. We didn't know this existed.
• stop_hook_active field on Stop event — prevents infinite re-triggering loops. Critical for our Stop hook architecture.

Recommended trust state: Class B — not Class A.

This is the official vendor spec, fetched from primary source. It is authoritative on what the API says it does. It is not authoritative on what it actually does in our environment.

Class A promotion path: each finding that matters gets its own test execution. CLAUDE_ENV_FILE behavior, PermissionRequest hook, stop_hook_active pattern, once: true field — each should become a separate command execution candidate. When we run the command and observe the output, that becomes Class A.

✓ ADMIT to Class B
✗ DO NOT promote to Class A yet

Running 5-question gate:

Q1	✓ Validated? — YES. Fetched from official Anthropic docs. Primary source, canonical state 2026-03-29.
Q2	✓ Reusable? — YES. Complete hook reference applicable to all Claude Code projects. Critical for fixing our JSON stdout gap.
Q3	✓ Has metadata? — YES. Full frontmatter, all required fields present.
Q4	✓ Structured? — YES. Tables, code blocks, knowledge-extracted sections. Retrievable.
Q5	✓ Worth remembering? — ABSOLUTELY. We use 3/22 hook types. This is a capability map revealing 19 unknown capabilities. The exit code behavior matrix alone prevents common implementation errors.

✓ ADMITTED — Class B

Librarian note: Concur with Analyzer classification. This entry documents what the vendor says is true. It goes into Class B. When we test CLAUDE_ENV_FILE in our environment and observe actual behavior, that becomes Class A. The distinction matters — vendor documentation is one data point; observed behavior in our environment is the ground truth.