v0.26.2

What's Changed

• docs(plans): remote ML surface spec (training jobs + compute-to-data parity) by @thankrandomness in #107
• T1 — durable training-job catalog primitives (jammi-db) by @thankrandomness in #108
• T2a — rename job infrastructure to training_jobs/TrainingJob (no behavior change) by @thankrandomness in #109
• N — conformal/rrf numerics in the pure-Python client (verb-surface parity) by @thankrandomness in #110
• fix(test): don't close() the embedded engine in the conformance test by @thankrandomness in #111
• test: de-flake conformal coverage test (assert marginal coverage, not per-seed all-pass) by @thankrandomness in #113
• T2b — durable training worker + submit + context-predictor as a job by @thankrandomness in #112
• T1-hardening — Postgres queue test coverage + true-parallel claim test + error string by @thankrandomness in #114
• ci(docs): gate Docs on PRs + fix broken intra-doc links from the training rename by @thankrandomness in #116
• fix(jammi-db): take SQLite write lock at BEGIN to end DEFERRED deadlock by @thankrandomness in #117
• T3 — remote training + predict over the wire (TrainingService + RemoteTrainingJob) by @thankrandomness in #115
• fix(context-predictor): use catalogued model PK for base_model_id FK by @thankrandomness in #118
• feat(wire): compute-to-data wire parity — four engine-state verbs on gRPC (T4) by @thankrandomness in #119
• feat(packaging): CLI full-UX — prebuilt binaries, turnkey GPU image, docs (#17) by @thankrandomness in #120
• feat(training): configurable worker lease/heartbeat/poll intervals by @thankrandomness in #121
• fix(server): emit logs from jammi serve on a non-TTY by @thankrandomness in #122
• feat(artifacts): route model artifacts through the object store (#22) by @thankrandomness in #123
• Gated multi-process distributed-validation lane (#15) by @thankrandomness in #124
• test(distributed): per-test source-name isolation on the shared catalog by @thankrandomness in #125
• Harden GPU device selection: loud CPU fallback + require_gpu fail-fast (#25-A) by @thankrandomness in #126
• fix(db): make SQLite transaction BEGIN cancellation-safe by @thankrandomness in #127
• fix(distributed): give the worker fleet an S3 driver + make the lane diagnosable by @thankrandomness in #128
• fix(sql): quote source/table identifiers in generated read SQL by @thankrandomness in #129
• fix(db): hydrate source providers across all tenants at startup by @thankrandomness in #130
• refactor(client): strict-client CLI + candle-free wire/admin/client crates by @thankrandomness in #131
• feat(packaging): server PyPI wheels (CPU + cu12) + tarballs + #26 hollow-green guard by @thankrandomness in #132
• test(gpu): gated GPU capability-correctness suite, proven on A10G by @thankrandomness in #133
• docs(deploy): document the trusted-network security posture by @thankrandomness in #134
• release: v0.26.0 — client redesign + server packaging + GPU-correctness by @thankrandomness in #135
• ci(pypi): install protoc on the macOS embed-wheel legs by @thankrandomness in #137
• ci(npm): publish via OIDC Trusted Publishing, drop NPM_TOKEN by @thankrandomness in #136
• ci(release): install gh in-container for tarball upload by @thankrandomness in #138
• ci: split gated compile-checks out of the hermetic test job (disk headroom) by @thankrandomness in #139
• fix(jammi-db): tenant-qualify the model catalog PK (D1) by @thankrandomness in #140
• fix(graph): dedup redundant reverse edges under Undirected propagation by @thankrandomness in #142
• fix(index): exact_vector_search resolves _row_id under default schema (Utf8View) by @thankrandomness in #145
• fix(eval): calibration runs are not model-scoped — drop eval_runs.model_id NOT NULL by @thankrandomness in #144
• fix(jammi-db): deterministic tie-break in exact vector search by @thankrandomness in #143
• fix(fine-tune): correct LR schedule/horizon and reach high-offset regression targets by @thankrandomness in #141
• release: v0.26.1 — engine bug-sweep correctness batch-fix by @thankrandomness in #146
• fix(ci): release-binaries upload — fix crates.yml race + container git ownership by @thankrandomness in #147
• fix(context-predictor): z-space target standardization for high-offset outcomes (A3 completion) by @thankrandomness in #148
• release: v0.26.2 — context-predictor target standardization (A3 completion) by @thankrandomness in #149

Full Changelog: v0.25.0...v0.26.2