feat(memory): implement TrajectoryRiskAccumulator and ImplicitConflictDetector

[bug-ops]

Summary

• TrajectoryRiskAccumulator (zeph-memory, feat(memory): implement TrajectoryRiskAccumulator — shadow memory safety guardrail #4372): per-session shadow memory stream that ingests per-turn audit signals with exponential temporal decay and gates tool execution when cumulative trajectory risk exceeds a configurable threshold. Addresses multi-turn attack vectors (tool-chain chaining, persistent indirect prompt injection) that per-turn ContentSanitizer controls cannot detect. Based on MAGE (arXiv:2605.03228).
• ImplicitConflictDetector (zeph-memory, feat(memory): implement ImplicitConflictDetector — write-time stale memory resolution #4373): write-time predicate similarity detection for APEX-MEM using Levenshtein distance, staging conflict candidates in implicit_conflict_candidates (migration 090), and annotating SYNAPSE ActivatedFact results with is_implicit_conflict/conflict_candidate_id. Addresses the STALE benchmark gap (55.2% frontier model accuracy on implicit conflicts).

Both features are opt-in with enabled = false default — zero overhead when disabled.

Changes

• crates/zeph-memory/src/shadow/mod.rs — new: TrajectoryRiskAccumulator with noop mode, decay, ring buffer, is_blocked()/should_escalate(), 8 unit tests
• crates/zeph-memory/src/graph/implicit_conflict.rs — new: ImplicitConflictDetector, normalized_levenshtein, detect_candidates, stage_candidates, annotate_conflicts, 12 unit tests including DB tests
• crates/zeph-db/migrations/sqlite/090_implicit_conflict_candidates.sql — new staging table
• crates/zeph-sanitizer/src/audit.rs — new: AuditSignalType, Severity enums
• crates/zeph-config/src/memory.rs — TrajectoryRiskAccumulatorConfig + [memory.shadow_memory]; ImplicitConflictConfig + [memory.graph.implicit_conflict]; non-negative weight validators
• crates/zeph-memory/src/graph/activation.rs — ActivatedFact extended with is_implicit_conflict: bool, conflict_candidate_id: Option<i64>
• crates/zeph-memory/src/graph/store/mod.rs — insert_or_supersede_with_conflict_detection wrapper
• crates/zeph-tools/src/executor.rs — ToolError::TrajectoryRiskExceeded { score, top_signals }

Test plan

• cargo +nightly fmt --check — clean
• cargo clippy --workspace --features desktop,ide,server,chat,pdf,scheduler -- -D warnings — clean
• cargo nextest run --workspace --features desktop,ide,server,chat,pdf,scheduler --lib --bins — 10149 passed
• cargo test --doc --workspace --features desktop,ide,server,chat,pdf,scheduler — 19 passed
• RUSTFLAGS="-D warnings" cargo check --workspace --all-targets --features desktop,ide,server,chat,pdf,scheduler --locked — clean
• Security audit: zero unsafe blocks, SQL injection safe, shadow isolation structural (no sqlx/qdrant in shadow module), noop bypass impossible
• Config non-negative weight validation (rejects NaN, inf, negative)
• halflife_turns = 0 clamped to 1 with tracing::warn!
• SQLite 999-bind-param guard in annotate_conflicts (chunks of 499)

Closes #4372
Closes #4373