feat: add --keep-models flag and QMD_KEEP_MODELS env var for MCP server

[bagg-anon]

Problem

When running qmd mcp or qmd mcp --http as a long-lived daemon, the current behaviour disposes model weights after 5 minutes of inactivity (disposeModelsOnInactivity: true hard-coded in createStore). This means the first query after an idle period has to reload the GGUF file from disk — which can take several seconds and causes a brief VRAM spike.

That trade-off makes sense for one-off CLI usage (you don't want to keep VRAM pinned after a single query), but not for a persistent MCP server where the whole point is low-latency responses.

The LlamaCppConfig already has disposeModelsOnInactivity: false as its default for exactly this reason (see the comment in llm.ts: "Keeping models loaded avoids repeated VRAM thrash"). We just had no way to opt into it from the outside.

Solution

Add a --keep-models flag (and QMD_KEEP_MODELS=1 env var) that keeps model weights in memory between queries while still releasing contexts (the lighter per-session VRAM objects) on inactivity.

Changes

File	Change
src/index.ts	StoreOptions gains keepModels?: boolean; createStore reads QMD_KEEP_MODELS env var and passes the flag through to LlamaCpp
src/mcp/server.ts	startMcpServer and startMcpHttpServer accept { keepModels } and forward it to createStore
src/cli/qmd.ts	qmd mcp gains --keep-models flag; it is also forwarded through the daemon spawn args; help text updated

Usage

# stdio transport (e.g. Claude Desktop, Cursor)
qmd mcp --keep-models

# HTTP foreground
qmd mcp --http --keep-models

# HTTP daemon (flag is forwarded to the child process)
qmd mcp --http --daemon --keep-models

# Via environment variable (handy for systemd units or Docker)
QMD_KEEP_MODELS=1 qmd mcp

Why not flip the default?

The 5-minute timeout + disposeModelsOnInactivity: true is the right default for CLI use — most people run qmd query once and move on, and holding VRAM indefinitely would surprise them. The flag is an explicit opt-in for operators who know they are running a server.

Testing

Verified by reading the logic path end-to-end:

• --keep-models → keepModels = true → createStore({ keepModels: true }) → disposeModelsOnInactivity: false → model weights stay loaded, contexts are still released on inactivity timer
• QMD_KEEP_MODELS=1 works the same way when the flag is absent
• Daemon spawn correctly appends --keep-models to child args so the flag survives qmd mcp --http --daemon --keep-models