feat: implement localpi runtime launcher

[dutifulbob]

Opened on behalf of Onur Solmaz (osolmaz).

Summary

Localpi was still mostly the old localagent wrapper.
This changes it into the local Pi launcher we wanted: llama-server by default, LM Studio as an explicit alternate runtime, default Pi tools, tool approval, token status, and no final-schema machinery.
It also installs the renamed localpi binary and keeps schema-constrained classifier output in caller tools such as localpager-agent.

What Changed

The public CLI now matches the renamed project.
The runtime layer now decides whether to manage llama-server, talk to LM Studio, or use a custom OpenAI-compatible endpoint.

• Renamed package metadata and binary from localagent to localpi.
• Removed --final-schema, --schema, the generated final_json extension, structured-output tests, and the example schema.
• Added model alias resolution with built-in Gemma aliases plus LOCALPI_MODELS_FILE overrides.
• Added managed llama-server start/reuse/status/stop support with pid and metadata under localpi state.
• Added explicit lmstudio and openai-compatible runtimes.
• Added generated Pi extensions for tool approval and token status.
• Added default tools: read,bash,edit,write,grep,find,ls.
• Updated README, runtime spec, structured-output note, and implementation plan.

Testing

I tested both the normal project checks and the actual local runtime path.
The smoke prompt reused the already-running Gemma 12B llama-server on port 18194; LM Studio reported no loaded models.

• npm run check
• node dist/src/cli/main.js --help
• node dist/src/cli/main.js --list
• node dist/src/cli/main.js --status
• node dist/src/cli/main.js --model gemma-12b -p "Reply with exactly: ok"
• node dist/src/cli/main.js --final-schema schema.json
• npm link
• localpi --list
• localpi --model gemma-12b -p "Reply with exactly: ok"

Risks

The main behavior change is intentional: localagent is no longer shipped by this package, and schema output is no longer part of localpi.
Older external workspace wrappers that still call localagent --final-schema need to move to localpager-agent if they are still used.

• localpi --stop only stops localpi-owned llama-server processes. It will not kill an unrelated server already running on the same port.
• LM Studio runtime support was checked through code and unit tests, but I did not load a model into LM Studio for smoke testing to avoid double-loading large models.

[dutifulbob]

Implemented the localpi runtime launcher end to end and pushed the final fixes through 0fe0727.\n\nValidation:\n- npm run check passes locally: format, lint, typecheck, 25 tests, build.\n- CI test job passes: https://github.com/dutifuldev/localpi/actions/runs/27062904972/job/79878669417\n- Smoke tested the built CLI against the already-running llama-server on 18194: node dist/src/cli/main.js --model auto -p 'Reply with exactly: ok' returned ok.\n- Confirmed LM Studio has no loaded models with lms ps before the smoke.\n- Confirmed removed schema mode fails cleanly: --final-schema exits with localpi's migration error.\n\nReview loop:\n- codex review found no P0/P1 issues.\n- Follow-up P2 findings around managed llama-server reuse, process ownership, model-id resolution, and LM Studio port warnings were fixed and covered by tests.