# Notarius v0.6.0

This release improves the reliability and ergonomics of unattended,
subprocess-driven D&D extraction pipelines.

## Summary

Notarius now gives models more precise, semantically useful correction guidance
when deterministic or model-backed validation rejects an otherwise structured
candidate. Semantic reconciliation retries identify candidates with compact,
request-local numbers, preserve valid candidates when a proposal cannot be
repaired, and report bounded process warnings when reconciliation is
incomplete. D&D extraction also canonicalizes safely resolvable reversed source
ranges and applies more consistent schema, prompt, and validator constraints,
improving successful operation with smaller models without weakening evidence
validation.

An optional LLM-backed combat-semantics validator can review whether scene
descriptions correctly consolidate combat and apply the `combat` kind. For
subprocess callers, command-line references now follow the pipeline-scoped
configuration model: a shared reference can be supplied once and automatically
reach every compatible selected target, while lane- and binding-specific forms
remain available for exceptions.

## Compatibility

- The meaning of an unqualified `--reference slot=path` or
  `--without-reference slot` selector has changed. It now applies to every
  selected pipeline target that declares the slot instead of requiring exactly
  one matching target. Review callers that relied on ambiguity rejection or a
  unique implicit target; the current selector contract is documented in the
  [CLI reference](../cli.md).
- The stage-wide `merge.slot` CLI reference shorthand has been removed. Use
  `lane.merge.slot` for an exact merge binding, `lane.slot` for all compatible
  bindings in one lane, or an unqualified `slot` for pipeline scope.
- Command-line file references cannot replace, remove, or coexist with a
  generated artifact handoff for the same concrete target and slot. Resolution
  reports the conflict so the caller can narrow or remove the selector.
- The private LLM-facing semantic-reconciliation contract now uses contiguous
  `candidate_number` values rather than application identities. This is not a
  durable or operator-configurable contract and requires no operator action.
- Durable D&D artifact schemas, the `notarius.run-result.v2` receipt,
  `notarius.warnings.v2`, and `notarius.diagnostics.v1` are unchanged from
  `v0.5.0`.

## Upgrade

1. Update subprocess commands to provide shared inputs once with
   `--reference slot=path`, and convert any stage-wide `merge.slot` selector to
   an explicit supported scope. The [subprocess consumer guide](../consumers/subprocess.md)
   and [complete D&D consumer guide](../consumers/dnd-pipeline.md) show the
   current invocation pattern.
2. Review CLI reference overrides that overlap generated same-run references;
   narrow or remove an external selector rather than attempting to replace the
   generated handoff.
3. Run `notarius config validate --config <path> --pipeline <id>` for every
   deployed pipeline configuration.
4. Optionally add
   `extract/dnd/scene-descriptions/combat_semantics` to the scene-description
   validator chain where the deployment wants LLM-backed combat-scene review;
   see the [configuration reference](../config.md).
5. Perform a representative run with the deployed model profile and inspect
   the machine-readable receipt, warnings, and diagnostics before promotion.

## Changes

- Added feedback-aware, module-requested retry support for normalizers while
  keeping retry budgets bounded and preserving attempt diagnostics.
- Reworked semantic reconciliation around compact, contiguous candidate
  numbers, explicit namespace guidance, complete proposal validation, and
  model-facing retry feedback that omits opaque internal identities.
- Preserved candidates separately when an exhausted semantic proposal cannot
  be safely applied, with a bounded process warning describing the fallback.
- Canonicalized safely resolvable reversed D&D source-reference endpoints
  before deterministic coverage validation and clarified the shared evidence
  prompt.
- Improved D&D schemas, prompts, and validators with closed-value constraints,
  contextual correction guidance, shared source-range coverage logic, and
  clearer item-holder transition rules.
- Added the optional D&D scene-description combat-semantics validator, shared
  combat policy assets, evaluation fixtures, and retry-path coverage.
- Made unqualified CLI references pipeline-scoped, added hierarchical lane and
  exact-binding selectors, defined deterministic override precedence, and
  documented first-class subprocess use.