MiniLang Compiler 1.2.16

MiniLang Compiler 1.2.16

This patch release improves the shared managed-heap runtime in both compiler
implementations. Both CLI version flags and compile-time MINILANG_VERSION
report 1.2.16. Standard-library and native-media ABIs are unchanged.

Memory management

• Add allocation-free gc_stat(index) diagnostics with 15 scalar counters.
  Invalid indices/types return void. Hot diagnostic accounting is omitted
  when application code does not reference the builtin, including indirectly.
  Individual concurrent reads are supported; several reads are not a snapshot.
• Adapt default collection thresholds to retained live block bytes, within
  fixed bounds. Explicit CLI/runtime limits and disabled periodic GC retain
  their previous behavior; allocation-failure collection remains available.
• Accelerate small single-thread allocations when the free list is empty,
  and avoid repeated unsuccessful free-list searches with an invalidated cache.
• Mark reference-free leaf objects without adding them to the tracing worklist.
  Reserve the 64-MiB worklist separately and initially commit only 64 KiB,
  growing on demand instead of committing the entire worklist at startup.
• With --heap-shrink, discard full interior dead pages while preserving
  allocation metadata, free-list links and live neighbors. Windows uses a
  discard hint; Linux uses MADV_DONTNEED. Top-trim accounting changes only
  after a successful OS operation; Linux decommit errors are propagated.
• Remove an unused recursive ML helper, correct runtime comments and verify
  Linux runtime label emission order.

This retains the shared heap, per-thread stacks/TLABs and stop-the-world
mark/sweep collector. It is not a moving, generational or concurrent collector,
nor a complete size-class allocator. Shared-object writes still require
appropriate synchronization.

Evaluation and trade-offs

The pre-version-stamp Windows and Linux/WSL A/B measurements show:

• 75–83% less time in fragmented, unsuccessful-fit allocation workloads.
• 44–46% less time with a large retained live graph and repeated allocations.
• 16–18% less time in the leaf-marking benchmark.
• Approximately 64 MiB less Windows commit charge for the small control,
  not 64 MiB less resident RAM.
• Linux interior-hole RSS falling from 66.00 to 2.06 MiB after collection
  with heap shrinking enabled. Windows' discard hint did not produce an
  equivalent immediate working-set reduction.

There is no universal speedup: small allocation churn takes 2.4% longer
on Windows and 6.8% longer on Linux, and reference-graph marking about 3%
longer. Threaded cases show no throughput gain. Adaptive GC keeps more
temporary heap in the retained-live case (80 to 96 MiB final committed heap,
about 10 MiB more peak RSS). Compiler binaries are approximately 1% larger.

See the evaluation and raw samples.
Those measurements preserve pre-version-stamp hashes. They are targeted
microbenchmarks, not new MiniSQL, MiniQuake or HollowKeep performance claims.

Validation

• Python full regression suite: 158 passed, 0 failed, 0 skipped.
• Self-hosted suite: 136 core cases passed, 0 failed, plus outer
  language/runtime, object-pipeline, codegen and smoke checks.
• Supplemental memory-policy, first-class diagnostics, page-purge, TLAB and
  safepoint runtime matrices pass on Windows and Linux/WSL.
• 96 memory-fixture images match byte-for-byte per target/configuration
  across Python, Windows ML and Linux ML hosts and both pipeline options.
• Version/runtime fixtures execute successfully with six-way parity per target;
  both CLI flags and compile-time version values report 1.2.16.
• Windows Python bootstrap and native selfbuild are byte-identical. Python and
  Windows ML also produce byte-identical Linux compiler images.
• A full Linux-native selfbuild passed before the version-only stamp; it was
  not repeated afterward. Release-stamped Linux compilation/runtime fixtures
  were rechecked.
• All 53 standard-library modules are byte-identical between repositories.
• Structural codegen and canonical Linux runtime-blob checks pass.
• MiniDoc compiler reference: 37 files, 3,287 symbols, 0 warnings.
• Extracted Windows/Linux binary packages pass version, standard-library import
  and managed-memory smoke checks; manifests and archive checksums are verified.

The memory parity manifest
and version matrix preserve
the release-stamped fixture hashes.

Python accepts the object-pipeline option but emits its monolithic equivalent.
The existing native cstr return-lowering exception to general byte parity
remains unchanged; see compatibility scope.

Release-stamped compiler checksums

Executable	Bytes	SHA-256
Windows mlc.exe	55,015,936	0BB6AF2112F74DFF0E4EEC0584434B2174F5ACA45C852018875F711A891C959A
Linux mlc	55,017,456	986044376AD1BC6E625E5424607C876C5471991045A3E616D27672EED4BFE1C5

Archive checksums are supplied separately as .sha256 release downloads.

Downloads

This Python release supplies GitHub source archives. Ready-to-run Windows and
Linux packages are available in the matching self-hosted release.