Changelog: upstream review ports

CHANGELOG.md

@@ -12,6 +12,24 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Security
+- **Pairing rejects out-of-order messages (`eap-noob`, `nvpair-cluster-manager`)**: A completion before the PIN was applied could reach Registered with an empty-key MAC, a key exchange before negotiation crashed either side, and the joiner accepted EAP-Success without a verified completion. A failed pairing message now also releases its pending session.
+- **Trust-store paths are confined (`nvpair-cluster-manager`)**: A peer identifier such as `../identity` in an endorsed roster entry could write a pin outside `trusted/`. Unsafe identifiers are refused.
+- **Bounded reads and fail-closed peers**: Node-info, model inventories, error-store ingest and remote progress frames are size-limited; a proxy whose peer pin has gone refuses the dial instead of falling back to an unpinned transport.
+
+### Fixed
+- **Uninstalling an engine keeps your models and settings (`nvpair-engine-manager`)**: LM Studio uninstall ran `rm -rf ~/.lmstudio` and oMLX `rm -rf ~/.omlx`. Each manifest now declares its model store, removal is confined and preserves it, and only the runtime is removed. PAIR uninstalls a vendor-installed engine only if it recorded installing it (engines installed before this release: use the vendor's uninstaller). oMLX no longer offers uninstall: its runtime is the oMLX app, which PAIR does not install.
+- **Stopping a leftover engine works on macOS (`nvpair-engine-manager`)**: The process-image lookup read `/proc`, so stopping an engine PAIR lost the handle to was always refused.
+- **Manual nodes survive worker restarts (`nvpair-ui-broker`, `desktop`)**: The broker owns `configs/manual-nodes.json` and replays it into every new manual-nodes worker, so Tailscale and VPN peers are no longer lost when the worker restarts or the terminal interface drives the broker.
+- **Jobs that share a number stay apart (`desktop`, `nvpair-tui`)**: Jobs are keyed by origin, engine, run and id, so an Ollama job and an oMLX job with the same number no longer share a card or a row.
+- **Workload frames are broadcast in order (`nvpair-workload-manager`)**: A removal can no longer overtake its upsert and resurrect a job on a peer, and a failed emit's retry is no longer swallowed as a duplicate.
+- **mDNS sends from UDP 5353**: Queries and responses use the RFC 6762 source port.
+- **The local node's inventory refreshes (`nvpair-node-scanner`)**: VRAM, CPU and memory in the discovery snapshot no longer stay at their startup values.
+- Also: IPv6 link-local pairing keeps a locally valid zone; LM Studio and oMLX chat actions and every engine's model load (locally and from a peer) wait up to 10 minutes for headers; engine health probes reuse connections; every GPU shows on a node card; the renderer CSP is tighter; a closed stdout no longer crashes the app; model search finds Ollama models newer than the bundled catalog.
+
+### Added
+- **Configurable proxy header timeout**: `--response-header-timeout` or `NVPAIR_PROXY_RESPONSE_HEADER_TIMEOUT` (default unchanged at 120 s) for requests that queue or wait on slow model loads.
+
 ## [0.8.0] - 2026-11-02
 
 Unsigned, unnotarized **macOS arm64 release** for this fork. Windows, Linux