fix: prevent global hooks from firing inside SDK-spawned Claude subprocesses

Discovered during Memoria Phase 4 first compile run: when compile.py
invokes claude_agent_sdk.query(), the spawned `claude` subprocess
inherits the global ~/.claude/settings.json hook config. Its
SessionEnd hook then fires when the subprocess wraps up, triggering
flush.py against today's daily log — polluting the log with compile
metadata and creating a soft recursion (every compile call also
generates a flush call).

flush.py already had this guard (CLAUDE_INVOKED_BY=memory_flush set
at module top before any SDK import). compile.py / query.py / lint.py
did not.

Add the same guard to the other three SDK call sites with
script-specific sentinel values:
  - compile.py → memoria_compile
  - query.py   → memoria_query
  - lint.py    → memoria_lint

The sentinel value doesn't matter — both session-end.py and
pre-compact.py check `if os.environ.get("CLAUDE_INVOKED_BY"): exit(0)`,
so any non-empty value short-circuits. Using distinct sentinels makes
diagnostics clearer if a hook trace ever shows it.

Verified: imports clean, all 29 acceptance tests still pass.

scripts/compile.py

@@ -13,9 +13,17 @@ Usage:
 
 from __future__ import annotations
 
+import os
+
+# Recursion guard — must be set BEFORE claude_agent_sdk imports anything that
+# spawns a Claude subprocess. Inner SessionEnd / PreCompact hooks check
+# CLAUDE_INVOKED_BY at startup and exit if set, preventing the global hooks
+# from triggering flush.py against this process's daily log (which would
+# pollute the log with compile metadata and could create a recursion loop).
+os.environ["CLAUDE_INVOKED_BY"] = "memoria_compile"
+
 import argparse
 import asyncio
-import os
 import re
 import sys
 from pathlib import Path