Release v0.0.83

[0.0.83] - 2026-07-31

Added

• resume --retry-exhausted: Also retry items that have used up their retry
  budget. Previously the only way to get such an item moving again was to edit
  the session state JSON by hand.

Fixed

These four defects compounded on a self-hosted migration where a WAF in front of
both instances intermittently refused requests: ~2,700 items failed permanently,
583 experiments were left empty on the destination without any error saying so,
and the summary misattributed the whole thing to feedback replay.

• Proxy/WAF 403s are retried instead of killing the item: LangSmith always
  returns JSON error bodies, so a 401/403 whose body is not JSON came from an
  intermediary (proxy, load balancer, WAF) that refused the request before it
  reached LangSmith. Those now raise a new retryable UpstreamRejectionError
  and go through the normal exponential backoff, and the error message says the
  response did not come from LangSmith. Genuine LangSmith permission failures
  still raise AuthenticationError and still fail fast in a single request, so
  a misconfigured key is not slowed down. The borrowed response body is
  sanitized (control characters and escape sequences stripped, whitespace
  collapsed, truncated) before it reaches the console or on-disk state, since it
  comes from a source outside our control.
• A failed source run query no longer reports success: when
  POST /runs/query against the source failed, migrate_runs_streaming logged
  it and broke out of the loop, leaving the failed-run counter at zero. The
  caller concluded the run stage had succeeded and advanced to feedback, so the
  experiment was created on the destination containing no runs and the only
  visible symptom was feedback replay incomplete (0/N migrated). The failure
  now records a run_query_failed issue and propagates, so the item is marked
  failed with the real error. A page cursor advances only after its buffered runs
  are written; failed writes retain the current cursor for a safe replay.
• One failed pass no longer exhausts an item's retry budget:
  update_item_status incremented attempts on every status change,
  including pending -> in_progress and -> completed. A single failing pass
  through an experiment burned four increments against a cap of three, making
  the item permanently unresumable. attempts now counts failures only.
• Complete experiments are no longer reported as failed: feedback replay
  counted every record it fetched as "found" but did not count records a
  previous pass had already replayed (which are skipped by fingerprint) as
  migrated. A fully migrated experiment therefore reported 0/N and was marked
  failed on every subsequent run, and could never reach a verified state.
  Already-replayed records now count toward completion, and feedback_verified
  is set only when every record is accounted for. Source feedback query failures
  propagate instead of verifying an empty or partial inventory. Genuine gaps
  (unmapped runs, failed creates) still report with a per-cause breakdown.