prismaquant 0.2.0

What's Changed

• Gracefully handle missing MTP tensors in index by @JDWarner in #1
• Enable LFM2.5-8B-A1B (Lfm2Moe): profile + generic per-expert→packed loader, vLLM-servable by @RobTand in #8
• Fix #4 (fla fast-path poisoning) + #6 (Gemma4 rotary init) + multi-layer-type rope forward by @RobTand in #10
• CI: run the test suite on push and PR; green main again by @RobTand in #18
• DSv4-Flash probe enablement by @tylerwagler in #13
• Fisher: normalize h_trace by the global calib token count by @tylerwagler in #14
• Footprint: per-tensor source-byte manifest; reject negative floors by @tylerwagler in #15
• Solver: feasible-only termination + solve memoization by @tylerwagler in #16
• Allocator: packed serving groups become first-class DP units by @tylerwagler in #17
• Allocator: bit-exact re-encode pricing, gated on an identity activation path by @tylerwagler in #20
• Allocator: opt-in gate_up/down role split, hard-gated on per-role serving-profile support by @tylerwagler in #21

New Contributors

• @JDWarner made their first contribution in #1
• @RobTand made their first contribution in #8
• @tylerwagler made their first contribution in #13

Full Changelog: https://github.com/RobTand/prismaquant/commits/v0.2.0