v0.9.5 — let a model propose, and keep the human gate

Added

• Added model backed truth extraction: recall/truth_extraction/, turning memo prose into
  structured, quoted claims behind a refusing validation ladder. Off unless
  RECALL_TRUTH_EXTRACTION=1, runs on the ingest path only, never the query path. The
  extraction engine is a port with two implementations, a deterministic rules reference and an
  OpenAI compatible model engine (pip install "recall-rag[extract]"); whatever an engine returns
  clears the same ladder, so a model gains no ability to skip a rung.
• Added recall extract run|show, which reads a corpus and writes nothing, and
  recall rewrite plan|apply|reject|verify, which declares reviewed claims in corpus
  frontmatter. recall rewrite apply is a dry run by default and requires --reviewer and
  --note as argparse requirements, so the named human gate fires before any code runs.
• Added the recall_rewrite_plan MCP tool, read only. There is deliberately no
  recall_rewrite_apply: the MCP client is the model, and a reviewer id it can type is a field
  rather than a person. recall_reasoning_proposals and recall reasoning proposals gain
  include_extracted, defaulting to off so existing behaviour is byte identical.
• Added recall extract run --cache PATH, a persistent SQLite extraction cache, so re-ingesting
  an unchanged memo does not re-pay the engine for it. Entries are keyed on engine identity,
  engine revision, prompt revision, the file, its body and the corpus names, so an answer
  produced under one engine is never served for another. A path that is not a usable cache is
  refused before any engine call, a corrupt row is a miss and is re-paid, and a failed write is
  counted and reported rather than discarding the files already extracted. --cache was briefly
  a boolean because an earlier version accepted a PATH and ignored it; the flag came back when
  the persistence did. See docs/EXTRACTION_CACHE_DESIGN.md.

Fixed

• Fixed recall extract run aborting a whole corpus on one filename that is not valid UTF-8.
  A POSIX name arrives as a lone surrogate through Path.glob's surrogateescape, and it raised
  twice: once hashing the cache key, which is computed for every file whether or not a cache is
  in use, and again printing the report, because reconfiguring stdout's encoding resets its
  error handler to strict. The first discarded every file already extracted; the second threw
  away a completed extraction at the last step, exiting 1 with empty output. Such a name is now
  reported with its bad bytes escaped.

Changed

• BREAKING: PROPOSAL_SCHEMA_VERSION moved from 1 to 2, which rewrites every ip_
  proposal id in existence, including the checked in results/reasoning_session3_proposals.json.
  Version 2 adds declares_validity and declares_status to ProposedRelation, because a
  document asserting something about ITSELF is not a relation between two documents and forcing
  it into references would put a false relation into an audit record. The bump is the point:
  an id minted under a vocabulary that could not express validity must not be mistaken for one
  minted under a vocabulary that can. Anyone holding stored ip_ ids must re-derive them.

Fixed

• MTRAG Task B and C generation no longer scores an answer that the token ceiling cut off.
  benchmarks/mtrag/generation.py sent --max-tokens (512 by default) and never read
  finish_reason, so a truncated completion was written to the submission and judged as if the
  system had produced it. It now raises CompletionTruncated, unretried because the same ceiling
  cuts every further attempt, and the existing per-task quarantine keeps the task out of the
  submission and in the failures log.

Added

• recall setup gains an optional reasoning arm step, asked after the entailment judge question
  and before the CLAUDE.md scaffold question. Answering yes writes four new environment
  variables: RECALL_REASONING, RECALL_REASONING_MODEL, RECALL_REASONING_BASE_URL, and
  RECALL_REASONING_API_KEY. Answering no writes RECALL_REASONING=0 and nothing else, so
  "switched off" and "never configured" stay distinguishable in .env. The shipped reasoning
  tools do not read these variables yet; this writes the settings for a port the reasoning arm
  will use once it is built. See
  docs/REASONING_MODEL_SELECTION_DESIGN.md.

Fixed

• Corrected recall setup's refusal message for an embedder whose vector width conflicts with a
  table that already holds data. It previously pointed at a remedy that failed identically to the
  original problem. It now stops and tells you to choose an embedder matching the existing
  schema, or point setup at a fresh table name or database.