[Bug]: preset `base` is not enforced - the preset may use any model

[r4victor]

Steps to reproduce

1. dstack apply -f preset.dstack.yml --fleet <fleet> with:

   type: preset
   name: my-preset
   base: Qwen/Qwen3.8-27B
   trials: 4
   
   # Agentic shape: long context, short turns, most of it cache-hittable, small-scale concurrency
   min_context_length: 262144
   max_ttft: 5000
   concurrency: 4
   input_tokens: 131072
   output_tokens: 512
   shared_prefix_tokens: 130048
   
   env:
     - HF_TOKEN
   

2. Compare base and model in ~/.dstack/presets/<id>/final_report.json.

Actual behaviour

base is not enforced. The agent may serve any repo it chooses, including a model from a different family, and the preset is still verified and reported as successful.

Verification checks only that the service advertises the requested name and that the report echoes the requested base:

dstack/src/dstack/_internal/cli/services/presets/verify.py

Lines 99 to 100 in 2f024c6

	if service.model is None or service.model.name != preset_configuration.model.api_model_name:
	raise CLIError("Claude final service model name does not match the requested model")

dstack/src/dstack/_internal/cli/services/presets/verify.py

Lines 142 to 146 in 2f024c6

	if configuration.model.allows_variant_selection:
	if report.base != configuration.model.api_model_name:
	raise CLIError("Claude final report base does not match the requested model")
	elif report.model != configuration.model.exact_repo:
	raise CLIError("Claude changed an exact model request")

Observed with base: Qwen/Qwen3.8-27B: the report recorded model: Qwen/Qwen3.5-27B-GPTQ-Int4, a different model generation, and the exported service ran vllm serve Qwen/Qwen3.5-27B-GPTQ-Int4 --served-model-name Qwen/Qwen3.8-27B. Whether a substitution happens depends on the agent's judgment about what counts as a compatible repo, so it does not reproduce on every run.

Expected behaviour

base should constrain which repos are acceptable.

dstack version

master at 2f024c6 (source checkout; dstack --version reports 0.0.0). macOS 24.0.0.

Additional information

I suspect agent will use a different model if it fails to satisfy the requirements. In my case, it failed to meet TTFT on A100 for 2 trials and switched the model generation.