feat: add Nemotron 3.5 Lightning and Muse Glimmer local models - #131

…ality

- Changed the GitHub repository for the llama.cpp backend from `AtomicBot-ai/atomic-llama-cpp-turboquant` to `AtomicBot-ai/atopmic-llama-cpp-turboquant-nightly`.
- Implemented a new auto-update feature for the managed local models, allowing the system to check for and download newer versions of the backend automatically.
- Updated the local models catalog to include the new Qwen 3.8 model, increasing the total number of models to 11.
- Added tests to ensure the correct functionality of the new auto-update feature and the integration of the Qwen 3.8 model.

This update improves the system's ability to maintain up-to-date backend resources and expands the available model options for users.

… tests

- Removed the outdated output.txt file.
- Updated the LOCAL_MODELS_CATALOG to include the new Muse model, increasing the total count of curated models to 12 (Qwen, Gemma, Nemotron, and Muse).
- Adjusted related tests to reflect the new model count and ensure unique IDs across the catalog.

This update enhances the model offerings and ensures the tests accurately represent the current state of the model catalog.

- Updated the GitHub repository references in `AGENTS.md`, `README.md`, and `backend-installer.ts` from `AtomicBot-ai/atopmic-llama-cpp-turboquant-nightly` to `AtomicBot-ai/atomic-llama-cpp-turboquant-nightly`.

This change ensures consistency in the documentation and codebase regarding the source of the llama.cpp backend.

- staged backend download + atomic swap (never wipe a working install)
- failed auto-update no longer aborts managed start
- timeout on the GitHub releases check
- update only on a genuinely newer release (no nightly downgrade/thrash)
- drop the dead nemotron-think profile discriminant
- stop advertising unwired ATEM/Harmony tool calling for Muse Glimmer
- rebase config migration onto v41 and refresh AGENTS.md

- autoStartIfReady no longer double-checks: startDaemon takes
  backendAlreadyChecked, the adopt branch keeps its own check
- surface a deferred update in TUI and CLI instead of staying silent
- correct the v41 migration narrative and the single-flight comment,
  both of which still described replaced behaviour

TUI auto-start ran the update check first, so the user got a rendered,
typeable prompt with no model behind it while a 27-39 MB download ran.
Start first; run one deferred pass afterwards, off the start path.

The deferred pass must not stop the daemon it just started:
hasOtherLiveSessions skips our own pid by design, so it would have
reported 'no other sessions' for the very model the user is talking to.
keepDaemonRunning makes that explicit and defers the swap to next start.

Also bound the asset download with a timeout on both paths - the signal
parameter existed but was never passed, so a stalled-open connection
pinned the update for the life of the process. CLI 'models start' keeps
updating before start: it is an explicit one-shot command.

The flag is on by default and drives a background download, but the
only way to change it was hand-editing config.json - the CLI form
rewrites the whole file. U flips it, the panel shows 'auto-update off'
next to the backend tag, and the footer advertises the key.