[Epic][TS Calls] Introduce explicit fallback policies and partial semantic models

Goal

Make unknown-call behavior explicit and evaluate its effect on symbolic exploration and concrete reproducibility.

Scope

• Use the current USVM and native JacoDB TypeScript frontend.
• Reuse one typed unknown-call dispatch boundary.
• Select an immutable set of semantic models independently from STOP_PATH or FRESH_SYMBOLIC_RETURN fallback.
• Support guarded normal/exceptional successors, state changes and aliases, with an explicit residual domain routed to fallback.
• Execute source models through the normal EtsIR interpreter; use intrinsics for engine primitives unavailable or inefficient in ordinary TypeScript.
• Observe decisions through TsInterpreterObserver. Persistent path history and stopped-state snapshots are not production requirements.
• Select a small justified model set using a real-project census and existing-model audit, then compare policy/model configurations with real TypeScript replay.

First-priority gate

#383 is P0 and blocks completion of #365/#366 and further integration using their old contract. Reuse the simplification already in PR #377/#380; do not create a duplicate implementation.

A model needs an ID, a declarative target and execution. Do not reintroduce separate backend/registration/descriptor layers, redundant implementation-kind/precision values, manually maintained model versions, or a generic plugin framework.

Work

• Completed foundation: #361, #362, #363, #364.
• Priority simplification gate: #383.
• Model integration: #365, #366.
• Independent census and audit: #367, #368. Their preparatory work may continue while the gate is resolved.
• Final policy/model experiment: #385. Add bounded family implementation prerequisites only when the census justifies them.
• Optional research-only frontier capture: #373, only when a concrete consumer needs more than existing decision events.

Definition of Done

• All relevant unknown calls use one documented boundary or have an explicit exclusion.
• Fallback semantics are consistent; opaque returns do not claim to model arbitrary side effects or exceptions.
• Built-in model applicability depends on sufficient target/receiver evidence, not a method name or one possible type.
• Unsupported domains retain their configured fallback; trusted models have focused guard, alias and exception tests.
• One useful intrinsic and one ordinary TypeScript model demonstrate the two implementation mechanisms.
• Decision observation does not change exploration or require generic core-state changes.
• A frozen model set and development/held-out evaluation in #385 measure real replay-confirmed targets on comparable budgets.
• Mandatory language semantics and correctness fixes remain identical between experimental profiles.
• Negative results and unsupported cases are reported.

PBT generation, hybrid scheduling, LLM/MCP, a complete standard library and speculative research-state APIs are outside this epic.

Begin with blocking issue #383 and the simplification in PRs #377/#380, then trace the TsInterpreterObserver and EtsIR interpreter contracts. Review the dependent work in #365, #366, #367, #368, and #385; completion requires the documented boundary, justified model set, focused behavior tests, and replay-based evaluation described in the definition of done.