ESS 0.56.0

Added

• Entity Runtime lowering targets entity-core 0.28.0. A String type's alphabet lowers to
  alphabet on every string field, argument and member it reaches, and nested alphabets lower to
  their intersection. <text>.count lowers to <path>.count, counted in Unicode scalar values.
  Still refused by name: an alphabet on a declared response field, a text literal outside its
  field's alphabet, or nested alphabets that share no character (AlphabetUnsupported), and a
  text length read through a quantifier element or a union payload (TextLengthUnsupported)
  (beyond10x/entity-runtime#54).

Changed

• Validation refuses an accepting when: branch or an external: branch declared before a branch
  the held state selects (when_subject:, when_subject_state:, when_state_changes:) in the
  same command where one request can satisfy both input guards, as ESS-COMMAND-004 naming both
  branches, with the reorder as the hint. An external: branch with no input guard can always
  overlap, and so can guards the finite prover does not separate (an open Integer or String,
  an Optional input). Where both guards held, the model interpreter and Entity Runtime answered
  the first declared branch and the Rust and Go targets the held-state branch, as the precedence
  order says. Declared held-state first, every consumer answers alike, and reordering changes no
  answer. A specification that declares them the other way round no longer validates; to fix
  it, declare the held-state branch before the accepting or external one
  (#486).
• Generated Rust type libraries (ess generate types, ess generate schema types-bundle) no
  longer force serde_json/arbitrary_precision onto every consumer. The manifest declares a
  default-on exact-numbers crate feature that enables it, and only where the realized types hold
  JSON numbers in serde_json::Number or serde_json::Value, or decode through a Value a union
  whose alternatives can hold a number. Depending on the library with default-features = false
  gives binary64 numbers. types-report.json names each affected value under the new
  rust_exact_numbers obligation, and the integer obligation now says it is exact only with
  that feature.
  Regenerating changes Cargo.toml and the Rust types-report.json, so committed generated output
  differs until it is regenerated (#483).
• ess generate writes ess-output-state/3: a settled .ess-output/state.json no longer records
  the output root's absolute path (root) or its device and inode (directory); only a checkpoint
  with a pending transaction carries them, for recovery. The reader accepts /1, /2 and /3. A
  write-mode generation that finds a settled /1 or /2 record rewrites it as /3 even when no
  file changes, so committed records change once; --check does not report that as drift and
  prints one warning naming root and directory. Releases before this one refuse a settled /3
  record before writing, in write and --check mode, with
  error: invalid output state: missing field 'root' (exit 1): a CI job pinned to an older ess
  fails from the first commit of a /3 record, so move such a pin together with the regeneration
  (#484).
• Generation no longer replaces an owned file whose bytes differ from what .ess-output records,
  in the folder that generated it as well as in a clone. It refuses before writing, naming the files
  and the re-enroll route: move them aside, remove .ess-output, run ess generate output adopt
  once per recorded owner, then regenerate. A missing owned file is still recreated.
  ess generate --check names such a file with that route, and the single-file --check routes
  refuse with it. A commit that updates generated files without their .ess-output record now
  refuses regeneration instead of being repaired (#484).
• A leftover .ess-output/state.next (never authoritative) is removed by every write-mode
  command that finds the root settled, and ess generate --check reports one; recovery and
  adoption close with an ess-output-state/3 checkpoint
  (#484).

Fixed

• Concurrent ess generate runs that create different new output roots under one directory, such
  as a shared $TMPDIR, no longer refuse each other as output ownership busy. A run locks only
  the root it creates exclusively, and its ancestors shared; two runs creating the same root still
  exclude each other (#485).