[ashhart]

Refs #1009 (the finding from 28 Oct). #2106 fixed it for Qwen3.8-27B. This finds where else it holds.

A row gives probability 0 to a key it must not see, so a masked key whose value is +Inf or NaN should add nothing to that row. In many prompt-attention kernels, 0 × Inf gives NaN instead. This adds a contract census of that case across every Metal family with its own prompt attention:

• one cache coordinate poisoned with +Inf or NaN;
• the rows that must not see it checked by class against an independent float64 formula;
• every finite reference output held to its existing bound;
• finite shape panels at 8 and 32 rows, five-repeat stability, and planted controls.

The findings are registered as known, as #2106 registered the 27B's before fixing it, so nothing new goes red on main. Flash Next is left out (its own line).

On an M5 Ultra, 45 panels are red and 37 green on main:

red: hidden rows see NaN	green
Gemma 4 (12B, 31B, E4B; global, sliding, wrap), Spark2.5 (full, sliding, wrap), MiniMax-M3 prompt, Llama (Nanbeige 3B and 33B), Nemotron prompt, LFM2 prompt (nax), GLM tensor direct (64 and 32 heads, padding, block split), Step5 tensor (+Inf only)	Qwen 3.5 small dense, GLM hybrid dense latent and generated indexed, GLM EXLR8 MLA, Kimi K3/K3.1 MLA, Step5 scalar, LFM3 on its PR head

Fixes follow family by family, the 27B's way: finite prompts run main's kernel unchanged, and the masked read sits only where a masked non-finite value can occur.

• Build: zig build tf-masked-values and zig build test-masked-values-host (macOS, outside the default install). The census also runs inside tf-kernel-contract.
• Checks:
  • test-masked-values-host: 148 of 148 tests;
  • a plain zig build: 255 of 255 steps;
  • zig fmt --check;
  • no new lean finding (metal_main.zig stays at 600 lines).