Declare native output defaults through existing typed dtype policy (#213)

[trissim]

Refs OpenHCSDev/openhcs#213; paired OpenHCS draft
OpenHCSDev/openhcs#217. Linnaeus owns this backend/dtype
integration. Stacked on PR1 (retained history/runtime-handle ownership), base
fix/durable-decorator-context-20260929; do not merge/install either without
coordinator acceptance. The PR diff adds only the dtype default and regressions.

Concrete defect / reproducer

ArrayBridge direct decorators default to PRESERVE_INPUT even when a callable
naturally produces floating measurements. Removing a BaSiC wrapper's own
astype(input.dtype) is insufficient: the outer dtype wrapper rescales/casts
its first output. Input uint16 plus output [-1.25, 12.5, 70000.5] loses values
unless an explicit native runtime config is passed. OpenHCS pipelines already
default to native via DtypeConfig, but direct calls are a distinct real entrypoint.

Existing owner extension

dtype_config_default accepts the existing DtypeConversionConfig owner. Its
runtime default and reflected signature use the same object; explicit
runtime dtype_config still overrides it through existing conversion runners.
No extra policy registry, special BaSiC name dispatch, wrapper bypass, module
global config mutation, fallback or new config schema. Framework/GPU wrapping
and undeclared preserve-input behavior are retained. OpenHCS declares its own
existing DtypeConfig native default on the BaSiC callable.

Evidence / scope

• Python3.14 AST parse: changed module + regression file pass, no imports.
• Focused Ruff E9/F63/F7/F82 and git diff --check pass.
• Four real NumPy wrapper regressions added: native fractional/negative/large
  range; explicit override with float sidecar retained; existing undeclared
  preserve behavior; untyped default rejected. Not executed yet under the
  serialized runtime schedule. This is a source checkpoint, not live readiness.
• No installed package/environment changed. Parent #151/#212 acceptance finished;
  Lovelace currently owns finite heavy slot, Confucius prepare-only next.

Focused catalog review: BOUND-2 existing typed config owns policy; IMPL-13 one
existing conversion mechanism handles both defaults and overrides; TIME-7 no
copied framework policy. New conversion modes still extend their existing runner
family, not new consumer branches. Source/AST/catalog only, no global NRA proof.

Remaining: focused regressions, exact paired OpenHCS direct and compiled/MCP fit
journey, retained history compatibility, current installed entrypoint acceptance
after reviewed integration. Keep draft and no issue-closing claim.

[trissim]

Parent integrated PR2 normally with current main409b1e0, published d6d92a7 without editing the original worktree. All four real callable dtype tests passed as part of the paired15-test OpenHCS/BaSiCPy source run: native fractions/negative/large output, explicit runtime override, unchanged undeclared preserve-input behavior, untyped default rejection. Actual source paths asserted, existing Python3.12 environment, no install/download. Original receipts in the persistent issue-batch/basicpy-parent-numeric-20260930 ledger. Parent owns the paired PR217 integration and installed/MCP acceptance separately; this is not a live activation claim.