v3.3.0 — Layered Memory & Lineage, Operational Hardening

Added

• Layered memory (L0→L3), ported from TencentDB-Agent-Memory's semantic pyramid: mem_nodes (immutable evidence: raw transcript lines, archived body revisions) + mem_lineage (derived-from/revision edges) give mem_items real traceability for the first time. mem_items.layer discriminates L1 atomic facts (deterministic hash-slugged, deduped), L2 daily scene digests (cumulative same-day merge with [[slug]] backlinks), and L3 a rolling persona (background + working rules, rebuilt from the newest L1 facts). All distillation is local/deterministic — no LLM calls.
• raios mem history <slug> and raios mem list --layer <n> — inspect a memory item's revision chain and filter by pyramid layer.
• raios sessions --canvas <session_id> — folds a session's session_events stream into a compact Mermaid flowchart; consecutive same-type events collapse into one node with a se:<id> back-reference to the full, untruncated payload — compression is never irreversible.
• raios usage now reports live Claude Pro/Max quota remaining. The statusLine script caches rate_limits.five_hour/seven_day usage percentages (from Claude Code's own stdin JSON) to ~/.claude/raios-usage-cache.json; raios usage reads that cache (with a 24h staleness cutoff) and shows 5h:XX% 7d:YY% remaining plus formatted reset times instead of a hardcoded "unknown".

Fixed

• mem_items.body unbounded growth: mem_upsert previously concatenated every write onto the same row forever. It now replaces the body and archives the previous version as an immutable mem_nodes revision — the full history is still recoverable via raios mem history, but the live row stays bounded.
• A second instance of the same bug, caught only by a whole-branch review: the 90-second periodic memory-sync thread re-scans the entire session transcript on every tick (fixed start timestamp), and was inserting a fresh, undeduplicated mem_nodes/mem_lineage row per matched fact on every pass — silently recreating unbounded growth one layer down. Fixed with content-addressed dedup on (project_key, kind, content) for l0_raw nodes (revision nodes are correctly exempt — each is a genuinely distinct snapshot).
• mem_upsert's archive-then-replace sequence (revision node + lineage edge + item update) now runs inside a single SQLite transaction instead of three unguarded autocommit statements.
• raios security / raios refactor output now discloses its own limitation inline ("pattern-based scan — a clean result is not proof of absence") instead of only in internal docs — both are regex/heuristic scanners, not semantic analysis.
• gen-context.config.json used an unrecognized customOutput key that sigmap silently ignored, so it never wrote SIGMAP.md (it was defaulting to .github/copilot-instructions.md instead). Corrected to the real output key.

Changed

• session_memory.rs (974 lines after the memory-layering work) split into a session_memory/ directory module — transcript_io.rs, heuristics.rs, distillation.rs, plus a thin mod.rs orchestrator. Pure move, no behavior change; full external call surface (auto_sync_agent_memory, collect_transcript, decision_lines_from_transcript, etc.) preserved.