Fix canonical FP32 gradients and BF16 branch graph lifetime

[CompleteDotTech]

Changes

Refs #3 and #4. This is a numerical/source repair and scoped evidence delivery; neither issue is closed.

• Make CUDA FP32 query dispatch depend on absolute positions, with fixed short eager padding, CPU-initialized RoPE bits, and coherent double shared K/V and LoRA accumulation.
• Charge materialized double GQA storage correctly and include the cache planner in CUDA resume identity.
• Release completed BF16 branch graph owners before optimizer allocation, with an actual CUDA lifetime regression and a failing old-source control.
• Publish opt-in FP32 HF/seam reproducers, portable numerical/mechanics/failure receipts, and preserve the original 45-feature matrix.
• Exclude generated TypeScript dependencies/build output from the source distribution.

Verification

• Final current-source CPU suite: 858 passed, 55 explicit skips (69.85 s); compilation, critical Ruff, whitespace, source-index audit and 45-feature/450-cell verifier pass.
• Independent source review found no blocking defect. New lifetime regression passes; proper old-source control fails; all 14 selected BF16 CUDA tests pass (counts overlap).
• Actual pinned-weight FP32: independent HF hidden result exact and all 96 Q/V adapter gradients pass; four short/seam cases plus 8191/8192 boundaries and a 10520-token state pass all 100 trainable gradients in each batched/serial/full comparison. Frozen tolerances remain unchanged.
• Public HF command and all four public short/seam cases pass on the frozen trained checkpoint.
• Noncheckpointed native rank-8 FP32 and BF16 each complete training, reload, same-run resume to step two, and reload. These are generated short mechanics, with independently initialized heads.
• Rebuilt wheel matches all 32 runtime modules and four license files. Source archive inspection found zero generated client dependency/build members.

Limits retained

The exact 32768-branch/65536-aggregate FP32 training resource probe OOMs during forward at the unchanged 70% allocator cap. Full-weight noncheckpointed FP32 resume also OOMs. The graph cleanup allows the same-init BF16 first optimizer step and reload, but resumed backward still OOMs. An allocator experiment is explicitly unsuccessful: expandable segments are unsupported on this Windows runtime. All failures and artifact/attempt histories are preserved.

Historical numerical/rank-8 receipts retain their original source hashes; the later graph cleanup has a separate source-bound receipt. This does not establish full-weight resume, installed serving for every mode, representative predictive quality, or Jev benefit. Overwatch remains archived by user direction.