[experiment] measure binding-symbol pruning (Tier-1 aliases vs Tier-2 containers)

[Fedr]

Experiment — not for merge

Measure the build-time and symbol-count impact of pruning Python-binding symbols via mrbind --ignore / --skip-mentions-of, to separate two hypotheses:

1. Tier-1 (typedef aliases): Vector_*_value_type, *_const_reference, *_iterator, etc. Per mrbind docs these become Python aliases (≈free to compile) — expected to declutter the namespace but not move build time, and --ignore may not even suppress auto-aliases.
2. Tier-2 (obscure/nested containers): std::vector<std::vector<…>>, std::vector<std::array<…>>, std::optional<std::variant<…>>, std::variant<…Polynomial…>, etc. — these are real py::bind_vector/container instantiations (the hot stl_bind.h cpp_functions in the -ftime-trace profile), so this is the actual build-time lever.

Method

A single same-runner step (vcpkg x64, Release, Clang 21) does 3 full bindings rebuilds with -B, appending flag lines to the shared scripts/mrbind/mrbind_flags.txt between stages:

• BASELINE (unmodified flags)
• +TIER1 (typedef-alias --ignore)
• +TIER2 (additionally the obscure-container --skip-mentions-of)

Each stage reports wall time and len(dir(mrmeshpy)) (symbol count via import). Same runner → clean deltas. Flags are edited at runtime only (nothing committed) — this is a measurement, not a change.

Because the flags live in mrbind_flags.txt (loaded by generate.mk for every platform/config), whichever ignores prove worthwhile would apply everywhere with a one-line edit.

CI scoped to linux-vcpkg only (other platforms disabled via labels).

[Fedr]

Closing — experiment complete (not for merge).

Measured the build-time / symbol-count impact of mrbind --ignore pruning (vcpkg x64, Clang 21, same-runner 3-stage -B rebuild):

stage	wall	mrmeshpy symbols
baseline	665.4 s	2793
+Tier-1 (typedef-alias --ignore)	666.3 s	2628 (−165)
+Tier-2 (nested-container --skip-mentions-of)	build failed	—

Findings:

• Tier-1 (value_type/reference/iterator/… aliases): --ignore does remove them (−165 symbols) but build time is unchanged (+0.9 s, noise). They're Python aliases, not compile cost → namespace cleanup only, no build-time benefit.
• Tier-2 (blanket --skip-mentions-of on nested containers) breaks the build: skipping a container type that's an aggregate-struct member leaves the struct's generated aggregate constructor referencing a now-unbound type (overload ambiguity std::vector(size_type) vs MR::Id's implicit →int). Carpet-pruning containers by regex cascades.

Conclusion: there's no free --ignore-regex win for the bindings build time. The cost is the bound container types + per-function cpp_functions; pruning them en masse breaks generated code. Real levers = surgical per-symbol API reduction (validate each, handle the aggregate cascade) or more vCPUs (the build is ~98% parallel → scales wall down with zero API loss).