Changelog

0.5.0 (2026-04-01)

Breaking change: complete architecture rewrite.

Removed pyproject.toml dependency — Pantry now uses importlib.metadata directly. Works everywhere: development, installed packages, Docker, notebooks.
Removed packaging dependency — zero external dependencies, pure stdlib.
Removed dependency groups — has_group() removed. Use has() directly.
Removed Pantry.discover() and Pantry.from_pyproject() — no longer needed.
On-demand probing — nothing happens at import pantry time. Packages are probed on first has(), get(), or [] call.
Added pantry.version(pkg) — returns installed version string.
report() accepts package names — pantry.report("numpy", "pandas").
Recommended pattern: if pantry.has("pkg"): import pkg — works with pipreqs and IDEs.
Source reduced from ~400 to ~300 lines (one file + bootstrap).

Lazy probing: external dependencies are no longer imported at import pantry time. Only metadata (installed? version?) is checked at startup. The actual import_module() happens on first get() or [] access. This makes import pantry fast even with many heavy optional dependencies (torch, pandas, etc.)
pantry.simulate_missing(*pkgs): context manager for testing fallback behavior. Temporarily hides packages so has(), get(), [], decorator, and report() see them as unavailable. State is restored on exit, even on exceptions.

License changed from Apache 2.0 to MIT
Promoted to Beta status
Enriched README and documentation (design philosophy, troubleshooting, examples)

pantry.lazy_import("dotted.path") — register own modules for deferred import
Breaks circular imports: register at top level, resolve on first pantry["path"] access
Two-step resolution: submodules via import_module, attributes via getattr
Results cached after first resolution
Separate from external dependency system (has, get, report unaffected)
Bridge toward PEP 690 (Lazy Imports)

First public release.