[Calls] Freeze explicit generic and union class input domains independently of outcomes #494

Part of #483.

Observed boundary

Priority: P1. 1 unbounded-generic first rejection; required by richer record values.

baloian--typescript-ds-lib HashUtils; original and engine-suited domain policies must stay distinct.

Implementation scope

Define a deterministic experiment policy for generic parameters, constraints and supported primitive/object unions across the entire class surface. Bind specialization/domain choices and their rationale in the manifest before pilot/profile results. Preserve one coherent choice across constructor, fields, members and returned observations. Support a declared bounded space rather than claiming arbitrary unbounded generic values.

Acceptance criteria

• HashUtils unbounded K receives either an explicitly frozen original-corpus specialization policy or an honest retained domain gap.
• Multiple supported kinds and optional/union cases use correct runtime-kind projection and native reconstruction.
• The previous engine-suited K=string choice cannot silently leak into the original population; revised scientific scope/domains are versioned and reported.
• Generic inherited types integrate with JacoDB [TS PBT] Search bounded generator choices while preserving dependent input structure #400 when required, without making it a prerequisite for every generic case.

Existing work to reuse or coordinate

• [TS PBT] Project property domains and preconditions into USVM #351
• [TS] Audit and formalize the type relations used by usvm-ts #359
• [TS] Preserve superclass type arguments in EtsIR jacodb#400

Evidence and delivery rules

Use the frozen v7 200-class ledger and the source-linked reproducer; do not overwrite sealed experiment output. Reuse existing loaders, domain projection, source instrumentation, candidate/replay contracts and bounded observations. Keep original class members and independently frozen domain policies. Run focused source/native checks for the changed contract and leave broad suites to CI; unsupported and newly exposed later gates remain recorded.

Local evidence: iccq2026-paper/experiments/class-coverage-round-robin-20261002/real-class-campaign-20261014/{original/selection-ledger-v7.json,V7_INFRASTRUCTURE_FRONTIER.md} and real-class-machine-diagnostic-20261014/revised-protocol-v8/README.md.