[ashhart]

Part of #1009. Stacked on #2322, which is stacked on #2308. Only the last commit is this PR's: "fix(lfm2): exclude masked nonfinite values from Metal prompt attention". It lands after #2322.

The fix: LFM2's prompt attention on Metal gets the same fix #2322 gives Nemotron. A non-finite value (Inf or NaN) in a masked key's V no longer reaches the output through a zero probability.

• A GPU scan over the newly written keys, staged tails included, picks between the original pipeline and an exceptional one, through indirect grids.
• Finite prompts run the same pipeline handles with the same grid as before.

Checks (Apple M5 Ultra):

• LFM2's kernel contract, both probability modes: 9 kernels pass, 7 planted controls are caught, and launch limits pass.
• Contract: masked non-finite values across every Metal prompt attention, the findings registered as known (refs #1009) #2308's census: 82 entries pass, and LFM2's two entries move from known to passing. 41 other-family findings stay known, and both controls are caught.
• Served, main against this branch: LiquidAI/LFM2.5-24B-A2B-MLX-4bit at 667408d4, warm base new new base, fresh servers, one and four streams, distinct chat and code prompts, greedy and seeded, with a 150 ms stagger. All 200 saved keys are equal across the four measured passes.

Prefill is level. Decode, aggregate tok/s:

Prompt	Temperature	Streams	Main	Fix	Change
chat	0	1	223.9	227.3	+1.5%
chat	0	4	477.1	475.1	-0.4%
chat	1	1	214.8	217.7	+1.4%
chat	1	4	420.2	421.6	+0.3%
code	0	1	247.8	252.8	+2.0%
code	0	4	498.8	506.2	+1.5%
code	1	1	238.8	239.1	+0.1%
code	1	4	419.2	421.1	+0.4%

These are within run-to-run spread, so no speedup is claimed. tests/contract/metal_main.zig goes from 600 to 547 lines. zig fmt --check passes.