chore(main): release 2.0.0

[github-actions]

🤖 I have created a release beep boop

2.0.0 (2026-09-10)

⚠ BREAKING CHANGES

• runtime: supporting taxis/cache/kernel APIs now return typed failures where construction or numerical success was previously unchecked. Shape/Layout/Tensor use checked element counts; DType uses checked byte counts; Tensor::from_cpu, Layout::from_parts, CacheLayout::new and FlatKvCache::new are fallible. The old alternate try_from_cpu entry is consolidated, public cache fields are private, fp32 softmax and unit normalization return Result, and embedding normalization retains the shared typed kernel cause. In-tree consumers and tests are migrated.
• text: decode processors, samplers and chain methods now return typed errors; policy construction is validated. Empty, NaN, positive-infinity and fully masked logits no longer silently select token zero. The in-tree consumer is migrated. The stable core trait/error surface is unchanged.
• runtime: the direct Rust library target changes from core to logismos_core. Cargo package identity and public facade/traits/errors are unchanged. Direct consumers using the old library-target import must migrate; existing project-qualified dependency aliases remain valid.
• runtime: cpu_f32::rms_norm returns kernels::Result<Vec>; transformers::Error::Kernel exposes a typed source. The unreleased Qwen35Weights::project_q8_0 is replaced by descriptor-driven project. Stable core traits, HIP FFI, and tolerance defaults are unchanged.

Features

• add agent-aware admission and artifact observations (#154) (f558245)
• add bounded native Qwen3 CPU reranking (#165) (e9b6537)
• add observation-bound model structure and q8 decoding (#156) (f6a21d7)
• bind native host accounting and recycled logits ownership (#191) (c75a1d9)
• bind native q8 execution to verified artifacts (#158) (109730d)
• bind shared native residency and prepared request ownership (#189) (e9a7238)
• cache: integrate execution-private paged KV (#183) (ea920d5)
• compose native gfx1100 full-attention layers (#185) (6d602a3)
• decoders: execute bounded native model chunks (#196) (4e28481)
• decoders: publish CPU sequence batches atomically (#198) (2d324f9)
• decoders: stage native single-sequence recurrent chunks (#195) (be44c41)
• derive native Qwen3 CPU retrieval requirements (#166) (2263ff4)
• embed: execute bounded native Qwen3 CPU embeddings (#164) (eab451b), closes #150
• establish agent-aware gfx1100 compute foundations (a0f4250)
• execute native gfx1100 hybrid main models (#187) (0c1e4a9)
• expose verified native text preparation (#176) (aeb6211)
• hermeneus: execute bounded native prompt chunks (#197) (08cfa58)
• inspect exact IQ4 artifacts and extend bounded emulation (#155) (fee3c55)
• kernels: add staged causal convolution HIP step (#181) (022436f)
• kernels: add staged grouped GDN decode step (#180) (4e7c2da)
• kernels: compose checked packed CPU prefill (#194) (e6dd128), closes #46 #150
• kernels: generalize native serialized-row GEMV (#182) (8af8165)
• kernels: integrate checked paged decode attention (#184) (52dde57)
• own shared native text execution and teardown (#192) (6eb1a12)
• prepare native recurrent hybrid composition (#186) (32fcf8b)
• prepare native text requests and Q8 GPU projection (#167) (bb1e746)
• preserve native and generated-output resource custody (#190) (6d4156f)
• refuse invalid native arithmetic before publication (#188) (981ce29)
• runtime: execute mixed-weight recurrent CPU paths (#159) (ec69f3d)
• runtime: execute native hybrid CPU model paths (#162) (e1b5203)
• sched: retire admissions without revoking peer workloads (#179) (9b1d7bd), closes #174
• text: add bounded native CPU generation (#163) (8fe4e6d)

Bug Fixes

• ci: repair release-checks caller permission (041dee5)
• runtime: enforce checked tensor and numerical boundaries (#193) (51b965d)

Documentation

• point planning guidance at private corpus (#157) (cb54330)

---

This PR was generated with Release Please. See documentation.

[forkwright]

Read-only follow-up for exact release head 0b0e10c: the rollup is now populated, but both jobs in https://github.com/forkwright/logismos/actions/runs/34332890580 failed because Cargo.lock needs updating under --locked. Boundary job 102405441319 failed while provisioning locked Rust sources; HIP job 102405441045 failed for the same lock mismatch before its code-object witness. This is separate from the immediate post-approval empty-rollup failure tracked at forkwright/.github#71. The --locked guard is exposing a release-branch consistency problem; no guard/workflow changes or release merge were attempted. Main PR #187 had a passing locked boundary/HIP candidate and exact merged-tree equality.

[forkwright]

The release-branch locked-provisioning failure recurs at 1dc74c1cdcd8b27dcbd0b3c893284d3144655189: HIP job 102429331380 fails before native compilation because Cargo.lock needs updating but --locked is required. This is separate from the post-approval rollup race and the passing #188 candidate. No release branch, lockfile, workflow or token was changed; removing --locked or merging this release is not part of the numerical-status delivery.