On the AST caching point, worth caching by structural hash of the parsed expression, not the source string. Copy paste with relative references produces different strings but identical AST shape, which hits a lot in financial-model-style workbooks where parallel columns share structure

Also worth a look: the "сalculation chain" docs in Microsot's OOXML SpreadsheetML spec describe how they serialize the dep order in xlsx files. Different problem (persistence vs runtime) but the data model is informative for what level of granularity ends up being practical

tidyverse + SQL + Terraform + dbt all have the same trick: the verbs literally ARE the operations, so reading the code IS reading the intent. Pandas, Ansible playbooks, hand-rolled ETL.. they all make you reconstruct intent from control flow. That works fine when writing is the bottleneck. It doesn't once review becomes the bottleneck.

Which is basically the AI era. Writing is free, reviewing isn't. Whichever grammar the reviewer can skim fastest wins the project regardless of what the model prefers to generate.

Kinda ironic given LLMs are way better at pandas than dplyr (training data). So now you've got this weird tension where the model wants Python and the reviewer wants R.