chore: release v2.50.22

[blisspixel]

This release packages the S0 rehearsal harness and a local study-prompt
repair. It adds no spend, mutation or remote authority, and it reports no
review, score or value result: the full 48-cell run and its human review are
still pending.

Added

• deepr eval expert-value-rehearsal plan|run|blind|bind-labels|workbook
  implements the local four-arm operational rehearsal and connects it to the
  existing value workbook.
  • run refuses before any work when the installed model digest,
    cloud-disabled status or a recorded $0 warm-up throughput floor fails.
    Each construct, maintain and answer phase runs in its own credential-free
    worker process with a newly materialized, verified source copy and its own
    data and cost roots, and every call is bound to the frozen digest.
  • Construction and maintenance never receive a question. Answers use
    identical instructions for every arm, read a private checkpoint copy hashed
    before and after, and are bound to their question digest.
  • Every cell ends answered, failed or blocked, including after launcher,
    copy or result-file failures; empty answers fail and length-stopped answers
    are flagged. Every attempt, including one interrupted or killed mid-call,
    is recorded at $0 in the worker and canonical ledgers.
  • operationally_complete requires every planned cell plus reconciled
    ledgers, unchanged consultation memory, checkpoints and canonical experts,
    and matching worker code and rendered sources. A check without evidence
    reports null rather than passing.
  • blind writes a reviewer packet without arm identity, masks harness
    markers an answer echoes while keeping both digests in the key, lists the
    label fields each case needs and any missing cells, and requires the key
    to live outside the packet directory. bind-labels checks exactly one
    label per answer against exact answer bytes.
  • run --resume continues an interrupted run only with byte-identical
    policy, plan and code, after re-verifying recorded cells, answers and
    checkpoints. Unfinished phases are kept as abandoned evidence and rerun
    under a new attempt id; an OS-level lock admits one writer per run root.
  • workbook assembles the strict deepr eval expert-value workbook from
    recorded execution and bound labels; the protocol attestation is filled
    only when the operator supplies their identity. Workbook v1 cannot
    represent failed trials, so a run with unanswered cells is refused.
  • No score, reviewer identity, winner or value claim is produced. See the
    design and
    practice review.
• The native Ollama investigation backend accepts a pinned integer seed and
  boolean think, and exposes the digest observed by its per-dispatch model
  attestation.

Fixed

• Local study prompts (expert make --local, expert study, formation) now
  place each source once inside the untrusted-content boundary. They had
  embedded the sanitizer result's Python repr, repeating every source three
  times with escaped newlines, which inflated prompts and could keep quoted
  anchors from matching the retained corpus. Existing study outputs are not
  rewritten; re-study to benefit.