Background API ignores the Context setting from Settings > Models (server is capped at 16K) #55

Hi, and thanks for TUFF. It's the best-performing local runtime I've tried (I've downloaded hundreds of models).

What I see

• Settings > Models > Memory (Auto off): Context = 64K, Expert cache = 16 slots (Gemma 4 26B-A4B, Ornith 1.5 9B / 35B-A3B).
• Using the Background API from a coding agent (OpenCode, openai-compatible, http://127.0.0.1:8080/v1), requests fail with effective prompt exceeds the configured context (context_length_exceeded) when the prompt is only about 16K tokens.
• In OpenCode's session database, the failing request followed a prompt of ~16.2K tokens, and another response ended with finish: length at exactly 16384 total tokens (prompt + generation). Nothing near 64K was ever reached.
• Turning Auto on shows the profiles Speed 8K / Balanced 16K / Context 200K. Context 200K makes the server very slow (the hint says it trades weight residency for context), and 64K is not available as a middle option for the server.

What I think is happening (from reading the source, so it may differ from the shipped build)
RoutedServerModelProvider.inferencePlan(for:) uses automaticPlan(for:) (hard-coded .balanced profile) or otherwise grows the model's runtime defaults to 16K/8K. It does not read the per-model memory profile that the app's own chat uses. ServerEngineRuntime then loads the session with installed.inferencePlan(for: descriptor).contextTokens.

Expected
The Background API / tuff serve should honor the context size chosen in Settings (or expose its own setting), as long as it fits the memory budget.

Suggested fix
A small opt-in override, e.g. an environment variable TUFF_SERVER_CONTEXT_TOKENS, used only when the working set still fits device.safeAppMemoryBudgetBytes (otherwise fall back to the computed plan). I wrote an untested patch (could not build it on Linux) that wraps inferencePlan this way; happy to turn it into a PR, or better, wire it to the Settings profile / the server settings file.

Environment
MacBook with M2 Pro, 16 GB RAM, macOS, TUFF 9.4.0), Gemma engine 1.0.20, OpenCode as the client.

Thanks!