llama.cpp's `llama-server` already speaks OpenAI chat-completions, so users
could already point Hermes at it via `--provider custom`. But "custom" means
they have to set OPENAI_BASE_URL by hand, the model picker doesn't list it,
and the dashboard has no way to surface the running server. This PR makes
llama.cpp a discoverable, zero-config backend.

What ships
==========

* `plugins/model-providers/llama-cpp/` — new ProviderProfile with default
  base_url `http://127.0.0.1:8088/v1`, aliases `llamacpp` / `llama.cpp` /
  `llama_cpp` / `llama-server`, and an offline-tolerant fetch_models override
  (returns None instead of raising when the local server is down).
* `hermes_cli/auth.py` — adds llama-cpp to PROVIDER_REGISTRY (modeled on
  the lmstudio entry: api_key auth_type with optional LLAMA_CPP_API_KEY +
  LLAMA_CPP_BASE_URL env vars). Removes the old llama.cpp/llamacpp/llama-cpp
  hardcoded aliases that pointed at `custom`, so the plugin's aliases win.
* `hermes_cli/models.py` — adds the same alias mappings to _PROVIDER_ALIASES
  so `--provider llama.cpp` resolves correctly through the CLI parser path.
* `hermes_cli/model_switch.py` — adds a probe-and-surface block in
  list_authenticated_providers, mirroring the existing lmstudio pattern.
  Three surfacing modes:
    1. Live probe: `${LLAMA_CPP_BASE_URL}/models` with a 300 ms cold-discovery
       timeout. If `llama-server` responds, the row appears with the loaded
       model. This is what makes the dashboard "magically" pick up a running
       server with no config.
    2. Hint mode: LLAMA_CPP_API_KEY or LLAMA_CPP_BASE_URL set, or current
       provider matches one of the aliases — 1.5 s timeout.
    3. Sticky current: when llama-cpp is the user's selected provider but
       the server is offline, the row still appears with current_model so
       the user doesn't lose access after restart.
  When no env vars, no current selection, and no live server, the row is
  not injected — keeps the picker tidy for non-llama.cpp users.
* `plugins/model-providers/custom/__init__.py` — drops the llamacpp / llama.cpp
  / llama-cpp aliases from the generic `custom` profile (they now belong to
  the dedicated provider).
* `scripts/start-llama-server.sh` — turnkey llama-server launcher whose
  default port (8088) lines up with the plugin's default base_url, so the
  end-to-end UX is just:
      ./scripts/start-llama-server.sh ~/models/foo.gguf
      hermes chat --provider llama-cpp
  Prints an alignment hint when PORT/HOST diverge from the plugin default.
* `tests/providers/test_llama_cpp_profile.py` — 12 tests covering plugin
  registration, alias resolution end-to-end through hermes_cli.auth,
  CANONICAL_PROVIDERS auto-injection, PROVIDER_REGISTRY entry shape, picker
  surfacing in three modes (current+offline, no-clutter, alias resolution),
  and the offline-graceful fetch_models override.
* `tests/providers/test_plugin_discovery.py` — bumped expected profile count
  33 → 34.
* `website/docs/guides/local-llamacpp-setup.md` — user-facing setup guide
  modeled on the existing local-ollama-setup.md.
* `website/docs/reference/environment-variables.md` — documents
  LLAMA_CPP_API_KEY / LLAMA_CPP_BASE_URL and adds llama-cpp to the
  HERMES_INFERENCE_PROVIDER accepted-values list.

Test plan
=========

  pytest tests/providers/                                # 90 passed
  pytest tests/providers/test_llama_cpp_profile.py -v    # 12 passed
  pytest tests/hermes_cli/test_model_switch_custom_providers.py \
         tests/hermes_cli/test_user_providers_model_switch.py \
         tests/hermes_cli/test_custom_provider_model_switch.py \
         tests/hermes_cli/test_api_key_providers.py \
         tests/hermes_cli/test_auth_provider_gate.py     # 221 passed

Tested on macOS 26.4 (arm64). The launcher uses
`nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || echo 8` so it works
on Linux + macOS + WSL2; not exercised on native Windows.

Notes
=====

* Existing PR NousResearch#19607 also adds `scripts/start-llama-server.sh`. The version
  in this PR supersedes that one — it's stripped of the Qwen-specific
  detection branches (this PR is intentionally generic-llama.cpp only) and
  reworded around the new `llama-cpp` provider's defaults. Whichever PR
  lands second will need a one-line conflict resolution.
* Does not include `tools/local_web_tools.py` — that's orthogonal web-search
  work and remains in NousResearch#19607.

llama.cpp's `llama-server` already speaks OpenAI chat-completions, so users
could already point Hermes at it via `--provider custom`. But "custom" means
they have to set OPENAI_BASE_URL by hand, the model picker doesn't list it,
and the dashboard has no way to surface the running server. This PR makes
llama.cpp a discoverable, zero-config backend.

What ships
==========

* `plugins/model-providers/llama-cpp/` — new ProviderProfile with default
  base_url `http://127.0.0.1:8088/v1`, aliases `llamacpp` / `llama.cpp` /
  `llama_cpp` / `llama-server`, and an offline-tolerant fetch_models override
  (returns None instead of raising when the local server is down).
* `hermes_cli/auth.py` — adds llama-cpp to PROVIDER_REGISTRY (modeled on
  the lmstudio entry: api_key auth_type with optional LLAMA_CPP_API_KEY +
  LLAMA_CPP_BASE_URL env vars). Removes the old llama.cpp/llamacpp/llama-cpp
  hardcoded aliases that pointed at `custom`, so the plugin's aliases win.
* `hermes_cli/models.py` — adds the same alias mappings to _PROVIDER_ALIASES
  so `--provider llama.cpp` resolves correctly through the CLI parser path.
* `hermes_cli/model_switch.py` — adds a probe-and-surface block in
  list_authenticated_providers, mirroring the existing lmstudio pattern.
  Three surfacing modes:
    1. Live probe: `${LLAMA_CPP_BASE_URL}/models` with a 300 ms cold-discovery
       timeout. If `llama-server` responds, the row appears with the loaded
       model. This is what makes the dashboard "magically" pick up a running
       server with no config.
    2. Hint mode: LLAMA_CPP_API_KEY or LLAMA_CPP_BASE_URL set, or current
       provider matches one of the aliases — 1.5 s timeout.
    3. Sticky current: when llama-cpp is the user's selected provider but
       the server is offline, the row still appears with current_model so
       the user doesn't lose access after restart.
  When no env vars, no current selection, and no live server, the row is
  not injected — keeps the picker tidy for non-llama.cpp users.
* `plugins/model-providers/custom/__init__.py` — drops the llamacpp / llama.cpp
  / llama-cpp aliases from the generic `custom` profile (they now belong to
  the dedicated provider).
* `scripts/start-llama-server.sh` — turnkey llama-server launcher whose
  default port (8088) lines up with the plugin's default base_url, so the
  end-to-end UX is just:
      ./scripts/start-llama-server.sh ~/models/foo.gguf
      hermes chat --provider llama-cpp
  Prints an alignment hint when PORT/HOST diverge from the plugin default.
* `tests/providers/test_llama_cpp_profile.py` — 12 tests covering plugin
  registration, alias resolution end-to-end through hermes_cli.auth,
  CANONICAL_PROVIDERS auto-injection, PROVIDER_REGISTRY entry shape, picker
  surfacing in three modes (current+offline, no-clutter, alias resolution),
  and the offline-graceful fetch_models override.
* `tests/providers/test_plugin_discovery.py` — bumped expected profile count
  33 → 34.
* `website/docs/guides/local-llamacpp-setup.md` — user-facing setup guide
  modeled on the existing local-ollama-setup.md.
* `website/docs/reference/environment-variables.md` — documents
  LLAMA_CPP_API_KEY / LLAMA_CPP_BASE_URL and adds llama-cpp to the
  HERMES_INFERENCE_PROVIDER accepted-values list.

Test plan
=========

  pytest tests/providers/                                # 90 passed
  pytest tests/providers/test_llama_cpp_profile.py -v    # 12 passed
  pytest tests/hermes_cli/test_model_switch_custom_providers.py \
         tests/hermes_cli/test_user_providers_model_switch.py \
         tests/hermes_cli/test_custom_provider_model_switch.py \
         tests/hermes_cli/test_api_key_providers.py \
         tests/hermes_cli/test_auth_provider_gate.py     # 221 passed

Tested on macOS 26.4 (arm64). The launcher uses
`nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || echo 8` so it works
on Linux + macOS + WSL2; not exercised on native Windows.

Notes
=====

* Existing PR NousResearch#19607 also adds `scripts/start-llama-server.sh`. The version
  in this PR supersedes that one — it's stripped of the Qwen-specific
  detection branches (this PR is intentionally generic-llama.cpp only) and
  reworded around the new `llama-cpp` provider's defaults. Whichever PR
  lands second will need a one-line conflict resolution.
* Does not include `tools/local_web_tools.py` — that's orthogonal web-search
  work and remains in NousResearch#19607.