[inureyes]

Summary

Gives the b10621 router the model cache source its four remaining deferred manifest entries needed, and implements --models-preset INI translation, closing what issue #1438 recorded as remaining after PR #1495.

What changed

• src/server/router_cache.rs (new): the router's cache source wraps the mlxcel model store (--model-store-root / MLXCEL_MODELS_DIR / MLXCEL_CACHE_DIR) behind a RouterDownloader trait, so tests drive the full download flow without the network; removal goes through the store's containment-checked deletion.
• src/server/router_models.rs: the pool reconciles cache, --models-dir, and preset sources with b10621's precedence; entries carry source, aliases, tags, hidden, per-model config, and a transient downloading state; start_download runs the fetch on a background task forwarding throttled download_progress SSE events; remove cancels in-flight downloads, unloads running instances, deletes from disk, and emits model_remove.
• src/server/router_presets.rs (new): b10621 INI grammar ([name] sections, [*] global cascade, long-option and LLAMA_ARG_* key spellings, version skipped, unknown keys a startup error), preset-only load-on-startup / stop-timeout / dedup-cache-models, and the CLI-wins overlay applied onto a per-model clone of the router's startup config, re-resolved through build_server_config so preset keys and CLI flags cannot drift.
• src/server/router_server.rs: POST /models validates (empty, exists, normalized repo id, synchronous metadata probe) then answers {"success": true} while the download runs; DELETE /models serves the full removal flow with upstream's exact refusal texts; the model object carries real source, can_remove, aliases, tags, status.preset, and download progress; dispatch resolves preset aliases and addresses the load path by the entry's real name (found live in the E2E: the alias path previously answered "not found").
• src/downloader/mod.rs: DownloadHooks (per-chunk progress, cooperative cancellation with a typed DownloadCancelled marker), download_repo_with_hooks, and probe_repo; the HEAD size pass also runs when a progress hook is installed so SSE events carry totals.
• src/server/startup.rs, src/server/cli_input.rs, both binaries: ServerStartupConfig derives Clone and carries model_store_root; router mode now also starts from --models-preset alone; --models-preset with a model argument fails startup loudly.
• Manifest compat/llama-server/b10621/router-models.json: POST /models and GET /models/sse flip to by_design with rationale and pinning tests (bare-name default-org expansion as policy; in-process payload residue as architectural, mirroring --models-dir), DELETE /models flips to supported, --models-preset flips to aliased with its refusal divergences recorded. docs/llama-server-compat.md updated to match.

Security

A download path plus a delete path is a new way to write and remove files. Names from HTTP resolve only through the registry or the store's sanitized <owner>/<name> composition; deletion re-asserts containment under the store root before removing anything; downloads write only under the cache root with the downloader's existing canonicalized-parent checks and symlink-safe tempfiles; the new routes sit behind the router's API-key middleware like every other management route.

Test plan

• cargo test --lib --features metal,accelerate --profile test-fast server::router (61 tests: discovery confinement, cache listing/collision precedence, download success/failure/cancel event sequences, remove refusals and disk deletion, preset parsing/overlay/dedup/alias routing, auth on all management routes)
• cargo test --profile test-fast --features metal,accelerate --bin mlxcel-server and --bin mlxcel (CLI arg structs changed)
• python3 scripts/ci/check_llama_compat_manifest.py
• cargo clippy --lib --tests --features metal,accelerate -- -D warnings, cargo fmt --all -- --check
• Real-checkpoint E2E on Apple Silicon: router started with a preset-defined smollm-135m-4bit (load-on-startup, [*] temp overlay, chat by alias), pre-seeded store snapshot deleted through DELETE /models, and mlx-community/SmolLM-135M-Instruct-4bit downloaded through POST /models over the real network (19 download_progress events with per-url {done, total}, download_finished, model_remove on delete, then a chat completion served from the downloaded cache entry)

Closes #1438