[AnsonDev42]

Interrupted ingestion could discard consumed inputs, concurrent SQL completion writers could reuse result cursors, and failures around batch submission or artifact retention could strand work or delete replay data. Preserve uncommitted ingestion slices, validate resume before side effects, serialize claims/cursors, and make cancellation irreversible.

Refactor submission into explicit claim/chunk/remote-create boundaries; prevent SDK create retries, release pre-create claims/capacity, propagate deadlines, and avoid request-file accumulation while capacity is unavailable. Preflight both generated export leaves before any writes; exclude private handoff records from source distributions. Harden artifact paths and atomic writes, normalize provider output errors, reject newer schemas, and release completed input/result buffers. Prepare patch 0.0.8, retaining schema version 6.

Proof

• Full configured pytest smoke: 345 passed, 9 skipped, 89.96% coverage; 85% gate unchanged. 57 additional parametrized cases, including 54 regressions demonstrated failing before their fixes.
• Targeted fake-provider/client suite: 79 passed. Isolated PostgreSQL 17 storage contracts: 45 passed, 4 backend-specific skips.
• Type checking, Ruff lint/format, strict documentation build and package build passed on the repaired source. The sdist excludes private handoff records; package contents were inspected and wheel source matched the checkout.
• Paid live checks: both OpenAI tests passed (3 requests, one attempt each; 48 input and 571 output tokens). Undiscounted inference-cost bound: $0.0002308, below the authorized $2 aggregate budget.
• Gemini inline/file live checks failed with HTTP 402 depleted prepayment credits before inference. Anthropic was skipped for missing credentials; Vertex was not run for missing configuration. No paid tests were repeated.

Boundaries

One execution owner per run remains required. Arbitrary iterators recover only in process; remote create ambiguity still needs verified operator resolution. Deadlines remain cooperative. See docs/design_docs/CODE_REVIEW_2026_11.md for findings, regression evidence and remaining limitations.

Independent review and fresh repair verification passed at 6b9bd58491384021296ed816fda76146e993578e: zero major findings, including 16 additional adversarial export checks. All eight checks passed on this exact head in CI run 43115222768. The two earlier orphaned workflow records still report queued but have zero jobs and no commit checks; the current PR/commit rollup contains only the eight successful replacement checks. No additional branch-protection/ruleset checks are configured.