$check_syntax() and $format() inherit stanc options meant for compilation

[jgabry]

Found during the design review for #1238. I asked Claude to write up the report below.

---

$check_syntax() and $format() silently inherit C++/stanc options that were
supplied for compilation, and $format() inherits the wrong ones.

What happens

Both read the constructor-options field directly:

• $check_syntax(), R/model.R:1117 — if (length(stanc_options) == 0 && !is.null(private$precompile_stanc_options_))
• $format(), R/model.R:1252 — stanc_options <- as.list(private$precompile_stanc_options_)

$check_syntax() at least falls back only when the caller supplied nothing.
$format() has no length check at all: it overwrites whatever the caller
passed with the constructor's options, unconditionally.

And because precompile_stanc_options_ is cleared by the first successful
compile (R/model.R:945), both methods see constructor options only, never ones
passed to $compile(), and see nothing at all once the model has been built.
So the current behaviour is not just leaky, it is inconsistent with itself
depending on whether a compile has happened yet.

Why it matters more after #1248

#1248 makes cpp_options and stanc_options persistent build configuration and
retires the precompile_* fields. The tempting migration is to repoint these two
reads at the new persistent field.

That would be the wrong fix. It would make build configuration permanently
leak into unrelated stanc operations — $format() would start reformatting under
whatever --O1 or optimisation flags were set for the last build.

Proposed contract

• $check_syntax() and $format() do not inherit persistent build stanc
  options. They use their own arguments.
• They do need the shared requirements that make the source parseable at all:
  include paths, and --allow-undefined when a user header is configured. Those
  are source configuration, not build configuration.
• pedantic stays an invocation-only convenience argument on each method.

Scope

Independent of #1238 — needs no build provenance. Behaviour change, so worth
doing in the 1.0 window alongside #1248 rather than after, since #1248 removes
the field these two currently read and something has to be decided at that point
regardless.

The $format() missing length check looks like a plain bug and is arguably
fixable on its own today.

[jgabry]

Closing — the premise no longer exists. Design note: #1254, implemented by #1256.

This issue is about $check_syntax() and $format() inheriting precompile_stanc_options_, options meant for compilation, and the open question was what they should inherit.

With deferred compilation removed and options one-shot, there is no precompile store: nothing accumulates stanc options between calls for these methods to inherit. Both operations also move to standalone functions — check_syntax_stan_file() and format_stan_file() — which take what they need as arguments and inherit nothing by construction.

So the question stops being askable rather than getting an answer. precompile_stanc_options_, precompile_include_paths_ and precompile_cpp_options_ all go with it (#1256).