How an unclosed tag fails
Badly. If <context> is never closed, everything after it — including your instructions and your output format — sits inside the context block as far as the model is concerned. It reads your commands as part of the document it was asked to analyse.
The symptom is a model that ignores instructions for no apparent reason, on a prompt that looks fine. It is one of the harder prompt bugs to spot by reading, and one of the easiest to detect mechanically.
Where it comes from
Assembled prompts. When sections are concatenated conditionally, a branch that adds an opening tag without its closing partner is easy to write and invisible in review.
Also from interpolated content: a retrieved document that itself contains angle-bracket text can look like a tag and unbalance the structure. Check the assembled prompt, not the template.
The adversarial case
A closing tag inside untrusted content is a documented injection technique. Text that includes </context> followed by instructions is attempting to escape its container and be read as a command.
Unexpected closing tags in a prompt built from user-supplied content are worth treating as suspicious rather than as a formatting mistake.