fix: clean up paged decode and SSM precision

[inureyes]

Issue #2230 left unused paged compatibility APIs, an unresolved model-owned storage default, and a CUDA bf16 SSM recurrence whose compiled preprocessing rounded dt' to bf16.

Changes

• Remove the dense-pointer and rotating paged compatibility surface, cache metadata, validators, and self-referential tests. Four pooled decode tests now use an independent explicit float32 attention oracle, retain the original absolute RMS bound, and add normalized RMS coverage.
• Cast both dt and dt_bias to float32 in the shared compiled preprocessing before Metal/CUDA/HIP dispatch. The pinned MLX promotion table had selected bf16 for float32 plus bf16; no tolerance changed.
• Commit a fail-closed paired/null storage driver, the complete 60-arm/180-cell GB10 sweep, exact raw output and provenance, and dated guidance. No cell cleared its repeated-paged null spread, so Gemma 3 and Llama 4 keep supports_paged_decode_backend() == true.

Measured evidence

For Granite seed 2069, normalized RMS/max fell from 1.957140831e-3 / 5.875451502e-3 to 2.890811210e-8 / 2.002812756e-7 at compiled dt', from 3.008783834e-3 / 1.075881018e-2 to 6.809246822e-10 / 4.717584582e-9 at decay, and from 2.532402801e-3 / 4.127457397e-2 to 2.838613609e-8 / 1.701009653e-6 at state.

The first Llama 4 512-token/concurrency-1 paged arm was an observed anomaly: 6.5 aggregate tok/s versus 11.9 dense and 11.8 repeated-paged, yielding +83.08% paired and +81.54% null deltas. Every record is retained; its cause is unproven, and the preregistered rule leaves the cell unresolved.

Validation

• Required removal search is empty; four pooled CUDA tests pass, and a temporary 1.01 output mutation made all four fail.
• Five focused CUDA SSM parity cases pass. Granite fused-kernel and graph paths produced identical 64/64 greedy IDs under deterministic SDPA.
• Workspace/all-target CUDA clippy, formatting, the three fast contracts, release build, and the full GPU-locked CUDA gate pass; the full gate reports zero failures.
• The measured analyzer validates exactly 60 arms and 180 successful cells; nine protocol fixtures, both artifact manifests, and the branch-wide whitespace check pass.
• Canonical English and Korean technical reports are included under TECHNICAL_REPORTS/ for PR fix: clean up paged decode and SSM precision #2270.

Closes #2230