v1.6.0

Added

• aval relevant — which decisions bear on what you are about to touch.
  resolve is exact, so a caller has to know a key's name before it can ask
  anything, and at the start of a task that is a chicken-and-egg problem: the
  way to learn that a decision governs a file is to already know the decision.
  The alternative was HEADS.md, which is every decision ever made. This ranks
  the vocabulary against --path (repeatable), --text, and --changed —
  what git reports modified, staged and untracked — and reports each ranked key
  with its verdict, from the same resolve everything else uses. So an
  undecided row says nobody decided it, which is the row that matters: it is
  where an agent left to itself invents an answer.

  Four signals, documented and weighted: BM25 over the query's words (1.0) and
  over the paths' own words (0.6), a record whose body names one of the paths —
  backticked, as a link, or as a glob (2.0 each, three at most), and what git
  says changed alongside the record (0.5 each, three at most). The tokenizer is
  frozen and the conformance battery compares the resulting order byte for
  byte. Nothing is trained, nothing is cached, nothing leaves the machine.

  It is a suggestion in the sense SEMANTICS §5.1 now defines: it resolves
  nothing, writes nothing, and is not part of what the corpus says is true. The
  exit code is always 0 — a ranking has no verdict, and "I found little" must
  not share a code with "I could not look" — except 2 for usage, including a
  --scope the corpus does not declare.

• --json carries dependencies, the ranked keys compacted for a router:
  key, state, exit, adr and unresolved, the last true for exactly
  undecided and contradiction. A dispatcher routes a change whose decisions
  are not settled to a person rather than to a worker without reading the rest
  of the payload.

• The MCP tool aval_relevant, with the same inputs and the same bytes.
  Its description carries the obligation the surface makes easy to lose: a
  model reading a ranked list treats the first row as the answer unless it is
  told, in the description, that the ranking resolves nothing.

• A relevance conformance battery, conformance/relevant.json, over a new
  relevance-sample corpus. It pins the two properties that make a ranking
  reviewable: the same corpus and query give the same order byte for byte, and
  a ranked key's undecided or contradiction verdict is reported as the
  resolver gives it rather than softened or dropped.

Changed

• aval-core gains relevance (a frozen tokenizer and a BM25 index) and
  Json: From<f64>, which writes a non-finite float as null rather than a
  literal no JSON reader accepts. Still no external dependencies.

---

What's Changed

• chore: add relais.toml policy by @fredericrous in #27
• feat: aval relevant, retrieval at the input edge by @fredericrous in #26
• chore: v1.6.0 by @fredericrous in #28

Full Changelog: v1.5.0...v1.6.0