[Nyarlathoteppppp]

Problem and result

For a DeepSeek V4.1 model, setup writes --vision off (or on) into the model's args (setup.py:8270); DeepSeek configs have no vision section. When the app builds the run config for a load or a tune, native-control.mjs drops --vision whenever cfg.vision is falsy, which is always the case for DeepSeek. It removes only the flag, so the engine gets ... --eos 1 off --dspark auto and exits with unknown option off (exit 2), as reported in #4507 and confirmed by a second user on Windows.

After the change, a valued --vision on|off pair is kept as written and only a bare --vision (the Qwen/serve flag with no value) is removed when picture reading is off. tools/quant_quality.py had the same token-only drop and gets the same rule.

Related discussion

Fixes #4507 (duplicate #4505 has the same cause).

Approach

I did not remove the pair (the issue's suggested expectation): the DeepSeek engine loads its vision tower by default (src/dsv41/ds4_model.hpp:65, bool vision = true), so dropping --vision off would load the tower the user or setup turned off and cost memory. Keeping the pair keeps exactly the tower choice setup wrote; for DeepSeek the args are the only place that choice lives (the app's picture toggle needs a vision section). The Qwen behaviour (bare --vision removed when pictures are off) is unchanged.

strata_cli/launch/runner.py:321 has a similar token filter but skips deepseek41 configs, so it is left alone. Archived scripts under bench/results/ with the same pattern are not touched.

Validation

Passed:

• node --test apps/host/tests/draft-layer.test.mjs (Node 26.0.0): 7 passed. The new test loads a deepseek41 fixture with --vision off, --vision on and a bare --vision, each with and without a following flag; it fails on 46c152a (1 failed, 6 passed) and passes with the fix. The existing "picture reading off sticks" and "picture reading on keeps --vision" tests still pass.
• python -m pytest tests/terminal/test_launch_vision_value.py: passes with the fix; 4 subtests fail on 46c152a (value left behind by quant_quality.engine_args).
• The 10 host test files that reference native-control.mjs (models-folder, packaged-control, native-two-models, source-server-env, qa1-downloads, parallel-default, user-args, drafting, draft-layer, component-updates): 99 of 107 pass with the fix; the same 8 tests fail on 46c152a (97 of 105 pass there).
• Environment: macOS 26 (Darwin 25.5) arm64, Apple M1 Pro, Python 3.14.7, base 46c152a.

Failed:
none observed.

Not run:

• a real DeepSeek V4.1 load in the app (no model or GPU here), Windows/Linux;
• the full host suite (node scripts/test.mjs): it did not finish here (no progress after 35 minutes; several files fail to load with ERR_MODULE_NOT_FOUND because this checkout has no apps/host dependencies installed).

Limits

The DeepSeek picture setting still cannot be changed from the app's toggle; that is outside this fix.