v0.1.0

The first release: a typed expression core, a vault of Markdown and
HyperMarkDown documents, pipelines over its cards, two named retrievals, and
graph traversal that records why each node is present.

Language

• A typed expression core — scalars, collection literals, lambdas, field access
  and pipelines — checked before it is evaluated.
• type declarations parsed and checked, so std/ is read rather than assumed:
  core, collections, doc, graph, knowledge and retrieval.
• Scalar as a type, Data as a union, and a collection vocabulary settled on
  order: a sequence has one, a set does not, and take needs elements that
  carry their own.
• : as the subtype operator, replacing <:.
• A collection literal evaluates at its checked element type, and an unrelated
  tree is held as Data.
• Hit and Ranking carry the retrieval that produced them — query, index,
  model, model revision, metric, and whether the search was approximate.

Retrieval and graph

• lexical, which matches shared spellings offline with nothing to build.
• semantic, which scores against vectors computed ahead of time;
  contrib/semantics/ produces the index and the
  binary only reads it.
• Which model fills that index is the vault's choice. [semantics.embedder] in
  hql.toml declares one of two providers — local, a checkpoint in the
  producer's own process, or http, an OpenAI-shaped /v1/embeddings endpoint
  covering a hosted provider and a model server on localhost alike — and every
  build flag overrides it field by field. A key is named as an environment
  variable and never written into the committed file, and http needs an
  explicit revision, because an endpoint offers nothing to derive the pin a
  score claims. The producer gains no dependency: http uses the standard
  library, and indexing through one needs no torch.
• Scoring checks the precondition it rests on. A dot product is cosine only for
  unit-length vectors, so the producer normalises whatever an embedder returned,
  the index refuses to be written holding a vector that is not unit-length, and
  the reader refuses an index declaring normalized = 0 or a metric it does not
  compute. Truncated embeddings are why this cannot be implicit: a prefix of a
  unit vector is not a unit vector.
• expand, which traverses outwards and records the reason each node is
  present, so a graph of matches and their neighbours does not claim that all of
  them matched.
• Links traverse as uplinks and downlinks.

The command line

• eval, check, run, repl, render, builtins and config. - reads
  standard input wherever a file is taken.
• --format json emits the type, the value and the report queue for another
  program.
• hql render --write runs the hql#eval blocks in a document and writes each
  answer beneath it as an hql#result block, replacing the previous one.
• Reporting yields a value and a queue: --report strict stops at the first
  error and is the default, --report collect reports every error findable. The
  mode comes from the query, then the vault's hql.toml, then $HQL_REPORT,
  then the default.
• Exit 0 on success, 1 on a language or I/O failure, 2 on a usage error.
  Diagnostics carry zero-based UTF-8 byte ranges and render with a line, a
  column and a caret.

Extensions and execution

• An extension mechanism with declared contracts, and a checked execution tree
  with structural step contracts.
• MapReduce pipeline stages with pluggable execution backends.

Corpora

• tests/fixtures/birds/: fifty-six encyclopedia articles with an embedding
  index committed beside them.
• examples/birds/: linked HyperMarkDown cards with
  runnable collection queries, including intentional errors for duplicate keys
  and non-orderable sorted-map keys.