feat(dev): add managed local model serving tool

Priority Level

Medium

Task Summary

Add a source-tree developer tool that deploys and manages local models through one vLLM process lifecycle. The tool should let contributors run Anonymizer against pinned generation and detection models on a GPU host or inside the supplied container without introducing a second deployment domain.

Technical Details & Implementation Plan

The implementation should:

• Compile pinned TOML profiles into normalized plans with task, model, vllm, and local sections.
• Provide compile, launch, probe, status, and stop commands.
• Support stock Hugging Face generation models and NVIDIA GLiNER or GLiNER2 through vLLM Factory.
• Preserve process identity with launch receipts, PID start markers, and process-group ownership checks.
• Probe task-specific capabilities rather than treating a listening port as readiness.
• Bound termination and cleanup so failed or stale processes cannot hang the tool indefinitely.
• Run the same compiler and local-process lifecycle on a GPU host and inside the supplied container.
• Document the deployment workflow and provide representative pinned profiles.

Acceptance criteria:

• Compilation is pure and produces deterministic, digest-protected plans.
• Plan digests are documented as accidental corruption detection, not authentication or semantic proof.
• Generation and GLiNER profiles share one lifecycle implementation.
• Capability probes validate chat completions or GLiNER dynamic labels, offsets, and scores as appropriate.
• Status and stop operations reject stale or reused process identities.
• Cleanup is bounded and reports whether it completed.
• Tests cover lifecycle invariants and behavior rather than profile inventory tables.
• Documentation covers both direct GPU-host deployment and the supplied container.
• A full Anonymizer dataset run succeeds with locally hosted GLiNER and a generation model, and both services clean up successfully.

Out of scope:

• Public package APIs for deployment.
• Remote compute orchestration or attachment to existing endpoints.
• Docker-specific compiler branches.
• Native Transformers GLiNER serving.
• Model-cache discovery.
• Cryptographic authentication of plans.

Implementation is tracked in PR #212.

Dependencies

vLLM Factory integration is pinned by the repository's local-model dependency group.

Start by reviewing PR #212 and the repository's local-model dependency group, since the implementation is already tracked there. Check how the source-tree tool should expose compile, launch, probe, status, and stop, then verify the acceptance criteria for deterministic plans, capability probes, bounded cleanup, and both GPU-host and supplied-container workflows.