aprender 0.65.2
aprender 0.65.2 — the parity-instrument release, fully published
0.65.0 was cut at 587ad0797 and its crates.io cascade stopped at 48 of 74 crates: five sibling dev-dependencies carried a version through the workspace alias, so aprender-core (needs aprender-data) could not be uploaded before aprender-test-lib (needs aprender-core). 0.65.1 (cut at 752f55346) fixed that and reached 64 of 74 crates before aprender-test-lib failed its publish verification (an include_str! of a file outside the crate). 0.65.2 supersedes both for every crate (74/74 on crates.io, drained 2026-09-04 23:20Z); v0.65.0 and v0.65.1 stay where they were cut.
Fixed since 0.65.0
- Packaging (PMAT-958, #2866).
aprender-test-libembedsscripts/perf-matrix.yamlthroughbuild.rsand a vendored copy that may never drift; the CB-510 guard now reports anyinclude_str!/include_bytes!in host-compiled code whose target escapes the crate. - Publish cycle (PMAT-955, #2864). The five dev-dependencies are path-only;
check_publish_preflight.shrule R6 refuses any sibling dev-dependency that carries a version (fixture on both polarities, mutation RED). - Drain log (PMAT-954, #2864).
cascade-drain.shkeeps theDEFERlines that say why a crate did not publish. - wgpu init deadlock (PMAT-952, #2862). The
shared_instanceinitializer no longer takesDEVICE_INIT_LOCK, whichGpuDevice::newalready holds while calling it (ABBA on a process's first initialization; the 0.65.0 release clean-room parked 49 threads on it). - stdio MCP transport (PMAT-953, #2863; PMAT-950, #2860). A server that answers and exits before reading the request keeps its response and its exit status instead of surfacing
write stdin: Broken pipe.
What 0.65.x is
PP-LLAMA-001 v3.1: the parity spec and the instrument it describes — perf_gate.sh with merge and release phases, paired/interleaved receipts with bootstrap statistics, the PP-26 batch-invariance witness, the PP-9 spend ledger with its scanner (PMAT-930..935), the F-9 publish preflight, and UNMEASURED cells with owners and expiries instead of numbers. No parity ratio is published in this release; every c>1 band is INVALID-CORRECTNESS until the witness passes on the release hosts (see docs/specifications/0.66-performance-parity-report.md for the 0.66 plan).
Known
CHANGELOG.mdhas no 0.65.1 or 0.65.2 entry: the claim-literal baseline is line-keyed and refuses the coordinate shift a changelog insert causes (PMAT-956); the entry lands with that guard fix.