llm() needs SOME max_tokens value on every call (a provider truncates a
response with no clear error otherwise – a codebook/data-check pass over
many columns at once is the case most likely to need more than the
built-in default). Rather than requiring every module call to repeat
params = list(max_tokens = ...) individually, set it once per session
here; llm() falls back to this whenever a caller's own params does not
already specify max_tokens (an explicit params$max_tokens on a single
call still wins, same precedence as every other params entry).
