Add generic attention key-measure execution support

[xmarre]

Implements the generic core contract and execution support required for heterogeneous MiniMax-H3 mixed-grid attention.

What this adds

• canonical attention_measure_v1 normalization/validation with positive rational per-key measure and semantic digesting;
• H3 mixed-layout validation and optional VDN external-sequence API-2 geometry cross-checking;
• provider capability binding with concrete provider/owner/generation/preprocess identity;
• an O(T) contiguous FP32 natural-log key-measure buffer; no O(T²) measure tensor is materialized;
• exact physical K-block coverage for every non-unit-measure key interval, merged with any existing exact sink;
• exact dense weighted SDPA and a bounded streaming dense oracle;
• weighted BlockSparseAttention execution for direct and H3 chunked routes;
• measure-aware sparse calibration identity while preserving the existing unweighted pool key unchanged;
• cache-hit hardening: cached plans reuse only immutable measure materialization, while the live H3 layout, optional API-2 external geometry, and concrete provider capability owner are revalidated on every use;
• optional VDN epilogue delegation only when the concrete attention forward is VDN-owned. API-2 mixed-grid geometry alone does not imply that VDN is installed; the native projection path remains valid without VDN.

The weighted route preserves all Q/K/V rows. Spatial carrier density is represented by adding log(measure_k) to attention logits. It does not implement representative-K/V deletion or downsampling.

Compatibility invariants

• Existing unweighted sparse-attention calibration identity remains (block_index, rows, uuids).
• Query-prefix exactness and key-block exactness remain independent.
• Calibration values (kmean / vscale) are not modified by the key measure; only weighted cache identity is measure-specific.
• A cached measure tensor is not accepted as proof of current runtime geometry or capability ownership.
• A VDN external epilogue is required only when the actual attention forward is the VDN forward; stale or foreign ownership fails closed.
• Existing serialized workflows without an attention_measure_v1 request retain their prior numerical semantics.

Validation

The current clean head 8406da920f9df19cbc14b76aef1bcd3dc4cf1119 is one implementation commit on current upstream master 1d48d9cf7bcecb6022a87b3cb13e0fb435bf9b8a.

Focused cache-hardening run 34560181433 completed 11 passed across the attention-measure contract suite and the new cache regression. The regression first proves a legitimate cache hit, then separately rejects a changed H3 layout, changed API-2 external geometry, and replacement of the registered capability owner.

The exact clean head also passes fork Python-lint run 34560729867: both Ruff and Pylint completed successfully.

Current companion structural validation was rerun from Spectrum run 34561055104, pinned to:

• Core 8406da920f9df19cbc14b76aef1bcd3dc4cf1119;
• Flow a6bf395133c8ce496217ca9194f3ab3f2b9181ef;
• VDN d5f158a1c1750d79b37cb6ef22b0f14e6a8cbad6.

The weighted Core-BSA / Flow / VDN / Spectrum targeted suites completed 43 passed, 1 skipped. The single skip is the separate Untwist-composition fixture, which was intentionally not installed in that targeted run; it is not part of the weighted Core/Flow/VDN contract being revalidated.

These are structural/CPU compatibility evidence only. They do not establish GPU numerical or decoded-media correctness.

Still required before production release:

• compiled CUDA numerical validation of the weighted chunked sparse path;
• real SM120 Sol-Attn validation through the provider companion;
• end-to-end schedule/receipt validation proving no additional H3 NFE and preserving the applicable 18 logical / 14 actual / 4 forecast schedule;
• decoded-media and performance validation.

Authoritative architecture: https://github.com/xmarre/MiniMax-H3-Flow-Aligned-Regenerate/blob/3afc6c63063f410afec756c42811a8cf1ad2ed77/docs/MIXED_GRID_MEASURE_ARCHITECTURE.md

Companion design PR: xmarre/MiniMax-H3-Flow-Aligned-Regenerate#29