summarizer: isolate from user settings (settingSources: []) so it doesn't fire third-party lifecycle hooks

[migliolirobertoeric-create]

Problem

summarizer.ts runs query() from @anthropic-ai/claude-agent-sdk, which spawns a claude
subprocess. That subprocess loads the user's global ~/.claude/settings.json and fires every
lifecycle hook the user has configured.

The existing #87 guard (EPISODIC_MEMORY_SUMMARIZER_GUARD + shouldSkipReentrantSync()) only
stops episodic-memory's own SessionStart → sync → summarizer recursion. It does nothing for
third-party hook-based tools, which also fire inside the summarizer subprocess and have no way
to know they're running in one.

Real-world failure with ai-memory: its
user-prompt-submit / stop / session-end hooks fired inside the summarizer subprocess and
POSTed the summarization meta-prompt (Context: This summary will be shown in a list... plus
the embedded conversation head) to the ai-memory server as if it were genuine user activity. On the
affected user's server that produced 100+ junk sessions/<uuid>.md pages, 140+ poisoned
cross-agent handoffs, and a corrupted "where did we leave off" handoff served at the next session
boot.

Root cause

The summarizer is an internal utility call, but it inherits the user's full settings (hooks +
MCP) as if it were a real interactive session.

Fix

Pass settingSources: [] in buildSummarizerQueryOptions — the SDK's documented isolation mode
(claude-agent-sdk Options docstring: "Pass [] to disable filesystem settings (SDK isolation
mode)"). The subprocess stops loading ~/.claude/settings.json, so no lifecycle hooks fire
inside it — third-party or our own. This also closes #87 at the root (no SessionStart hook →
no re-entrant sync), making the EPISODIC_MEMORY_SUMMARIZER_GUARD env flag belt-and-suspenders
rather than load-bearing.

    persistSession: false,
    settingSources: [],   // isolate: internal call, must not fire the user's hooks/MCP

Why it's safe

Summarization needs only the model + auth (provided via env in getApiEnv()); it does not need
CLAUDE.md, MCP servers, or hooks. Auth (subscription / API key) is independent of
settingSources. The #87 env guard stays in place as defense in depth.

Testing

• npm test: this change adds no failures — verified identical results with and without the
  patch (206 passed). test/summarizer-options.test.ts still passes (it asserts individual
  option properties, not strict object shape). The one failure on my machine
  (test/show.test.ts:186, toMatch(/9\/19\/2025|2025-09-19/)) is a pre-existing,
  locale-dependent date-format assertion that fails under a non-US LANG; unrelated to this change.
• Suggested addition, in the spirit of test/sync-cli-reentrancy.test.ts: point a temp
  CLAUDE_CONFIG_DIR at a settings.json whose SessionStart hook touches a sentinel file, run
  one summarization, and assert the sentinel was not created.

Build

src/ + committed dist/ → npm run build and commit both (per the project CLAUDE.md).