feat(memory): add MemGuard type-aware retrieval composition

[bug-ops]

Summary

• Adds a retrieval-only, fetch-time gate on ContextAssembler::schedule_context_fetchers so a turn can compose evidence from only the functionally relevant memory type(s), instead of always fetching every source unconditionally.
• New FunctionalType enum (episodic / user_fact / behavioral_rule / reasoning_strategy / cross_session_summary / graph_fact) maps onto the existing per-collection/per-table memory sources — no new storage, no write-path change, category payload left dead/untouched.
• Five of six gated fetchers respect the active type set; fetch_corrections (behavioral rules / past corrections) stays unconditionally composed as a documented safety-critical invariant.
• [memory.type_aware_compose] config section, enabled = false by default — a byte-for-byte no-op identical to pre-change behavior. Optional intent_scoped widening reuses the existing heuristic memory router; no new LLM call in v1.
• Full integration: config/default.toml advisory block, --init wizard prompts, --migrate-config step, CHANGELOG entry.

Background

Code-grounded audit found Zeph already had functional-type isolation at the storage layer (six Qdrant collections, several dedicated SQLite tables) but no equivalent selectivity at retrieval time — every turn composed all memory types regardless of need. The design went through two adversarial critique rounds before implementation: round 1 (critical) corrected an initial premise that assumed a single undifferentiated vector pool; round 2 (significant) dropped an underspecified new trait method in favor of gating the existing per-slot fetchers directly, and corrected the user-fact source mapping. See .local/specs/064-memguard-type-aware-memory-retrieval/spec.md for the full spec and revision history (not included in this diff — .local/ is gitignored).

Closes #6086

Test plan

• cargo +nightly fmt --check
• cargo clippy --profile ci --workspace --all-targets --features "desktop,ide,server,chat,pdf,scheduler,testing" -- -D warnings
• cargo nextest run --config-file .github/nextest.toml --workspace --features "desktop,ide,server,chat,pdf,scheduler" --lib --bins (13456+ passed)
• cargo test --doc on all touched crates
• Rustdoc gate (RUSTFLAGS="-D warnings" RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links") clean, full workspace
• Dedicated regression test for the fetch-splitting fix (fetch_document_rag must survive Episodic exclusion)
• SC#4 test asserting actual PreparedContext type-exclusion at the gather() level (not just fetcher count)
• Independent security audit: zero findings, corrections path confirmed always-on, config parsing confirmed fail-closed on unknown type strings
• Independent performance validation: gate confirmed genuinely zero-cost when disabled, excluded fetchers confirmed never scheduled (not just filtered post-hoc)