sim-lib-lang-python

Embedded, capability-scoped, agent-authored Python scripting over SIM values. It is deliberately not a CPython replacement. Thin, bounded Python core evaluation enters through codec/python; the profile directly interprets the stable lowering and composes shared runtime organs, numbers, managed arena, and optional tracing collection.

PYTHON_FIDELITY reports syntax, lowering, direct evaluation, object/control, module/library, boundedness, and expected gaps separately. PYTHON_EVIDENCE_CASES publishes frozen call, cycle, exception, generator, matching, import, and capability-refusal cases. CPython 3.14.6 supplied offline expected values for simple differential cases only; it is never linked, imported, spawned, or used as a fallback. There is no bytecode, compiler IR, optimizer, project graph, or private language organ.

Capability-gated eval, exec, and imports

Dynamic Python is safer than ambient CPython evaluation by construction. DynamicPython sends both eval and exec through the installed source codec and the canonical diminished read-eval broker. The caller must supply a trusted ReadPolicy, already hold every required capability, and choose the smaller capability set visible to decoded code. Source text cannot mint authority.

PythonModulePolicy likewise resolves only through a caller-supplied Dir, then delegates decoding, failure/cycle caching, live exports, and receipts to the shared namespace module lifecycle. It never searches host paths.

python_library_manifest() derives present members from the checked Python matrix. Host-facing names such as open, os, subprocess, and socket are explicit absences; they never fall back to CPython or ambient host services. Structural cases compose ordered Shape checks and case-local captures. Guards run only after a shape accepts, and unsupported Python pattern forms remain a no-match gap instead of acquiring a second matcher.

PythonObjectSpace supplies checked class construction, C3 MRO, Python data/non-data descriptor precedence, bound methods, and super while leaving storage and interception in sim-lib-dispatch. PythonGenerator projects start/send/throw/close over sim-lib-control's bounded resumable frame; checked exceptions, groups, chaining, and synchronous context cleanup add no scheduler.

Tracing collection is the default constructor. PythonHeap::retaining is an explicit supported alternative and reports its strong-cycle leak gap.

The public PYTHON_OBJECT_CONTROL_GAPS list is the precise fail-closed boundary: custom metaclasses, post-construction descriptor-protocol mutation, weak-reference callbacks/proxies, __del__ resurrection, and async scheduling remain unclaimed.