feat: audit log, usage tracking UI, OpenAI orchestrator compaction, onboarding + docs

Tool audit log:
- Every orchestrator tool call logged to home/{user}/tool_audit/YYYY-MM-DD.jsonl
- Files panel sidebar: audit log group (collapsed), date-linked read-only table
- Admin endpoints: /api/audit/files, /api/audit/day, /api/audit/recent, /api/audit/stats
- Engine and model name recorded per entry

OpenAI orchestrator improvements:
- Context budget enforcement: 75% of model context_k (min 16k)
- Message compaction: truncates old tool results when approaching budget
- max_rounds respected per model config (intersected with server cap)

OpenRouter onboarding (setup.html, onboarding.py, app.js, settings.html):
- Step 3 of 3: /setup/model with curated model picker
- Chat banner for users on server-default model (informational, not alarmist)
- Settings quick-link card; /setup/model works standalone for existing users

Model registry + session store:
- set_role_config / get_role_config for per-role tool lists and system_append
- session_store: session rename, session name backfill endpoint

UI updates (app.js, index.html, style.css, local_llm.html):
- Role toggle in context panel
- Off-the-record mode
- Agent notes read-only viewer
- OPERATIONS.md loaded at T2+ in context

Documentation:
- HELP.md: full tool table, per-role tool sets, Agent Notes, usage tracking
- TOOLS.md: Agent Notes section, count corrected to 44
- ARCH__SYSTEM.md, ARCH__BACKENDS.md, MASTER.md updated to match reality
- CLAUDE.md: onboarding flow, documentation philosophy sections
- README.md: stack in practice, DeepSeek TUI mention, architecture diagram updated
- TODO__Agents.md: onboarding task completed with deviation notes

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

CLAUDE.md

@@ -146,8 +146,8 @@ http://localhost:8000/docs
 - Tools are registered in `cortex/tools/__init__.py` as both Gemini FunctionDeclarations and Python callables
 
 ### Context / Memory
-- `context_loader.py` assembles Inara's system prompt from `inara/` files based on tier (1–3)
-- Tier 1 = minimal (identity only); Tier 2 = standard (+ memory + user profile); Tier 3 = full
+- `context_loader.py` assembles Inara's system prompt from `inara/` files based on tier (1–4)
+- Tier 1 = minimal (identity only); Tier 2 = standard (+ memory + user profile); Tier 3 = + last 2 sessions; Tier 4 = + last 7 sessions
 - Memory files are written by the distiller or manually — do not delete them
 
 ### Security / Safety
@@ -160,6 +160,31 @@ http://localhost:8000/docs
 - Passwords are bcrypt-hashed and stored in `home/{username}/auth.json` — never in `.env` or the DB
 - Invite tokens are one-time-use, 72-hour expiry, stored in `home/{username}/invite.json`
 
+### Onboarding Flow
+New users follow a three-step setup before reaching the chat:
+1. `GET /setup/{token}` → password form → `POST /setup/{token}` sets password + session cookie
+2. `GET /setup/persona` → persona creation form → `POST /setup/persona` bootstraps persona directory
+3. `GET /setup/model` → OpenRouter quick-connect → `POST /setup/model` saves host + model + role assignment
+
+Step 3 is optional (skip link goes straight to `/{user}/{persona}`). `/setup/model` also works
+standalone (accessible from Settings) for existing users who haven't configured a model.
+
+All in `cortex/routers/onboarding.py`. Model writes use `model_registry.py`: `save_host()`,
+`save_model()`, `set_role(username, "chat", "primary", model_id)`.
+
+### Documentation Philosophy
+Cortex is a no-black-box system. Docs must match reality — at all times.
+
+- **Docs first:** When planning significant changes, update `TODO__Agents.md` and the relevant
+  `ARCH__*.md` to describe the intended design *before* implementing. This creates a spec to
+  implement against.
+- **Verify after:** Once implementation is complete, re-read the pre-written docs and confirm
+  they match what was actually built. Update anything that drifted.
+- **HELP.md is a user contract:** It describes what users can do. Never let it describe
+  features that don't exist or omit features that do.
+- **CLAUDE.md + ARCH__*.md are the developer contract:** Update them as the architecture evolves.
+- **Stale docs are bugs.** If you notice drift, fix it before moving on.
+
 ---
 
 ## Adding a New Tool
@@ -212,19 +237,23 @@ clearly asked for a directory to be unblocked.
 
 ---
 
-## Current State (2026-04-28)
+## Current State (2026-05-06)
 
 Cortex is running and stable. All channels are live:
 
 | Channel | Status | Notes |
 |---|---|---|
-| Web UI | ✅ Live | `https://cortex.dgrzone.com` |
+| Web UI | ✅ Live | `https://cortex.dgrzone.com` — PWA-installable |
 | Nextcloud Talk | ✅ Live | HMAC-signed webhook, async reply |
 | Google Chat | ✅ Live | Workspace Add-on, `hostAppDataAction` response format |
-| Local backend | ✅ Live | Open WebUI/Ollama, per-user multi-model config |
-| Orchestrator | ✅ Live | Gemini API tool loop → Claude response; ⚡ toggle in UI |
+| Local backend | ✅ Live | Open WebUI/Ollama on scott_gaming, per-user multi-model config |
+| Gemini orchestrator | ✅ Live | Gemini API tool loop → Claude response; ⚡ toggle in UI |
+| Local orchestrator | ✅ Live | OpenAI-compatible ReAct loop; fires when orchestrator role → local model |
+| Tool audit log | ✅ Live | Every tool call logged to `home/{user}/tool_audit/YYYY-MM-DD.jsonl` |
+| Token usage tracking | ✅ Live | Per-user `home/{user}/usage.json`; summary in Settings |
+| Web push | ✅ Live | VAPID push notifications; `web_push` tool; subscribe via ☰ menu |
 
-Active users: scott (inara, developer), holly (tina), brian (wintermute)
+Active users: scott (inara), holly (tina), brian (wintermute)
 
 **40 orchestrator tools:** web_search, http_fetch,
 file_read/list/write, shell_exec, claude_allow_dir,