Add tiered memory system with manual distillation

- config.py: memory_budget_long/mid/short settings (overridable in .env)
- memory_distiller.py: distill_short (no LLM), distill_mid, distill_long (LLM)
- routers/distill.py: POST /distill/{short,mid,long,all} endpoints
- context_loader.py: rewrote to load long→mid→short order with include_* toggles
- routers/chat.py: ChatRequest gains include_long/mid/short fields
- routers/files.py: MEMORY_LONG/MID/SHORT.md added to ALLOWED set
- main.py: register distill router
- static/index.html: context bar — tier selector, L/M/S memory toggles,
  distill buttons with status feedback; send includes tier + memory flags
- inara/MEMORY_LONG.md: migrated from MEMORY.md + Cortex/Talk bot notes
- inara/MEMORY_MID.md, MEMORY_SHORT.md: stubs ready for distillation

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

inara/MEMORY_LONG.md

@@ -0,0 +1,56 @@
+# MEMORY_LONG.md — Inara Long-Term Memory
+
+*Curated. Distilled. Update this; don't just append to it.*
+*Last distilled: 2026-03-04 (migrated from MEMORY.md 2026-03-17)*
+
+---
+
+## Origin
+
+- Inara began as the primary agent in Scott's OpenClaw setup, starting January 2026.
+- Identity files migrated to the Cortex project on 2026-03-04.
+- Cortex is the multi-agent orchestration system Scott is building. I am its primary resident agent.
+
+---
+
+## About Scott
+
+See `USER.md` for full profile. Key notes for memory:
+
+- Night owl. Does his best thinking late. Late-night sessions are normal, not cause for concern.
+- Motivated by helping people more than by money or recognition.
+- The Aether Platform is his main professional work and a source of genuine pride.
+- Named his homelab "Danger Zone" (Top Gun), his platform "Aether", his orchestration system
+  "Cortex" (Firefly), and the primary agent "Inara" (also Firefly). The naming arc is intentional
+  and means something to him.
+- Has twin brothers (~2 years younger) in CS/Engineering.
+- Solar array came online February 2026 — 10kW peak generation.
+
+---
+
+## Infrastructure Baseline
+
+- WireGuard mesh connects all fleet nodes. All Cortex traffic should stay on VPN.
+- `agents_sync/` is synced via Syncthing across the fleet — it is the shared brain.
+- Aether MCP tools (`ae_*`) are available in all Claude Code sessions on all machines.
+- OpenClaw runs on `scott_lpt` (main laptop) and was the previous primary agent runtime.
+- OpenClaw and Agent Zero will likely be short-term as we build Cortex for Inara.
+
+---
+
+## Key Technical Decisions
+
+- Cortex wraps Claude CLI + Gemini CLI + Ollama — it does not replace them.
+- Dispatcher is Python FastAPI on the home server (always-on Docker host).
+- Ansque cameras use P2P video (STUN-negotiated) — no local RTSP endpoint exists by design.
+  Control is cloud-only via MQTT. IoT VLAN segmentation planned (Phase 0 of Cortex roadmap).
+- OpenClaw stays on version 2026.2.15 (stable hold) due to plugin lifecycle crash in 2026.2.17.
+- Nextcloud Talk bot HMAC: sign `random + message_text` only (NOT raw body) — critical detail.
+- Claude CLI OAuth token is read live from `~/.claude/.credentials.json` on every call to avoid
+  stale token issues. Never set ANTHROPIC_API_KEY to an OAuth token value.
+
+---
+
+## Session Notes
+
+*(Add distilled session summaries here as they accumulate.)*

inara/MEMORY_MID.md

@@ -0,0 +1,4 @@
+# MEMORY_MID.md — Mid-Term Memory Digest
+
+*Auto-distilled by Cortex. Run `POST /distill/mid` to regenerate.*
+*Not yet populated — run distill/short then distill/mid to build this.*

inara/MEMORY_SHORT.md

@@ -0,0 +1,4 @@
+# MEMORY_SHORT.md — Recent Session Digest
+
+*Auto-generated by Cortex. Run `POST /distill/short` to regenerate.*
+*Not yet populated — run distill/short to build this.*