B7: Apple Silicon worker support (v0.19.0)

gpu.py: darwin branch via system_profiler SPDisplaysDataType for GPU
model/VRAM on macOS, falls back to rocm-smi on Linux.
hardware.py: _get_vram_darwin() parses system_profiler output.
node_agent/agent.py: get_load() reports VRAM on darwin via
system_profiler.
tests: 5 new gpu tests (linux/darwin/absent), 3 new hardware tests
(darwin assessment + VRAM parsing).

AGENTS.md

@@ -60,7 +60,7 @@ Refactored from single-file (`app.py`) into modules under project root:
 | `search.py` | SearXNG integration, perplexity scoring, refusal detection |
 | `rag.py` | Qdrant vector search + system prompt assembly + chunk_text() helper |
 | `eviction.py` | Score-based RAG eviction engine |
-| `gpu.py` | AMD GPU stats via `rocm-smi` |
+| `gpu.py` | GPU stats — `rocm-smi` (AMD/Linux) or `system_profiler` (Apple Silicon/macOS) |
 | `triage.py` | Phi-4-mini-based query classification + cluster node selection |
 | `cluster.py` | Cluster node registry, event log, coordinator election, ping/pong, model swap handlers |
 | `amqp.py` | AMQP connection manager — connect, disconnect, publish, subscribe, auto-reconnect |
@@ -135,6 +135,7 @@ All streaming endpoints yield `data: {json}\n\n`. Key shapes:
 ## Work State
 
 ### Completed this session
+- **B7 (v0.19.0)** — Apple Silicon worker support. `gpu.py` now detects `sys.platform == "darwin"` and parses `system_profiler SPDisplaysDataType` for GPU model/VRAM instead of `rocm-smi`. `hardware.py` has darwin branch via `_get_vram_darwin()`. `node_agent/agent.py` reports VRAM on macOS via `system_profiler`. 5 new tests cover linux/darwin gpu paths, 3 new hardware tests cover darwin assessment + VRAM parsing.
 - **Scroll-position fighting** — `scrollToTop()` now respects `_userScrolledAway` flag (100px threshold), skips auto-scroll when user is reading older content. `resetScrollLock()` called on new messages.
 - **401 error cascade** — `SESSION_TIMEOUT_SECONDS` bumped 90→3600 (1 hour). All 10 unprotected `authFetch` calls wrapped in try/catch.
 - **Token counter** — removed localStorage persistence; resets to 0 on page refresh.
@@ -156,13 +157,12 @@ All streaming endpoints yield `data: {json}\n\n`. Key shapes:
 - B4 — RAG Corpus Management UI
 - B5 — default model auto-pull on first start
 - B6 — waterfall direction toggle
-- B7 — Apple Silicon worker support (gpu.py Metal, hardware.py Darwin)
 - B8 — **Encryption & PHI readiness** — spec out encryption at rest (SQLCipher for caic.db, Qdrant payload encryption) and in-transit (TLS for inference, AMQP, RAG). Per-user auth, audit logging, log sanitizer, data lifecycle. Document the "personal LAN HIPAA gap."
-  - **Preferred approach: Private Chat mode** — a toggle that skips DB persistence, memory/RAG injection, and content logging entirely. Zero stored data = zero data to protect. Simpler, more robust, less code to audit than full encryption. Design this as the primary PHI path before reaching for crypto.
+  - **Preferred approach: Private Chat mode** — a toggle that skips DB persistence, memory/RAG injection, content logging, and **external SearXNG searching** entirely. Zero stored data, zero external queries = zero data to protect. Simpler, more robust, less code to audit than full encryption. Design this as the primary PHI path before reaching for crypto.
   - **In-transit still needs TLS** — Private Chat eliminates at-rest risk, but AMQP (RabbitMQ) and inference (llama-server) traffic between coordinator and workers is still plaintext on the wire. TLS termination on each node is the lightweight fix (self-signed CA, nginx sidecar or RabbitMQ TLS).
 
 ### Key config values (current)
-- `VERSION = "v0.18.0"` in `config.py`
+- `VERSION = "v0.19.0"` in `config.py`
 - `SESSION_TIMEOUT_SECONDS = 3600`
 - `DEFAULT_MODEL = "qwen2.5-7b-instruct"`
 - `LLAMA_SERVER_BASE = "http://192.168.50.108:8081"`