fix(playground): pull the speculative tier - quantized framing degrades output

[ronaldtse]

Pulls the speculative vocalization option introduced yesterday. New measurement (interscript-ml RESULTS.md, 2026-09-12) found two disqualifiers on quantized artifacts:

1. Decode framing changes the decode. Dynamic-int8 ONNX graphs compute activation quantization scales per fed tensor, so single-step and batched decode paths produce materially different numerics on the same machine. The batched-verifier path measured a word-short output on a canonical row ("امُ عليكم" vs "السلام عليكم") - a quality degradation, not byte noise.
2. Acceptance was framing-relative. The 0.9886 probe acceptance was measured under uniform plain-path framing; under the runtime's real framing it measured 0.4614. And int4 CPU kernels are slower than int8 here, so there was no speed story either.

The SpeculativeModel API stays in interscript@5.4.x as measurement infrastructure with the constraint documented; the playground serves only the verified tiers. All 255 tests green.