bulk-write-hardening.md

@objectstack/core	patch
@objectstack/objectql	patch
@objectstack/rest	patch
@objectstack/metadata-protocol	patch

fix: harden the bulk-write path — retries, idempotency, contracts, and summary visibility (#3147–#3152)

Six reliability fixes to the batched seed/import + engine.insert(array) path introduced by the #2678 bulk-write rework:

• #3151 bulkWrite validates that writeBatch returns one record per input row (a short/long/non-array return is degraded per-row, not backfilled as phantom success); engine.insert(array) likewise rejects a short driver bulkCreate return instead of padding afterInsert with undefined.
• #3150 wraps the two remaining un-retried write points (seed writeRecord/resolveDeferredUpdates, import's no-createManyData fallback) in withTransientRetry; defaultIsTransientError short-circuits definitive logical errors to non-transient.
• #3148 import resolveRef flushes pending creates on a same-object miss so a later row can reference an earlier same-file CREATE, and no longer negatively caches a miss.
• #3149 threads an attempt counter through bulkWrite; seed rechecks by externalId and import by matchFields before re-writing, so a commit-then-lost-response retry cannot duplicate a batch.
• #3147 recomputeSummaries retries transient failures and, on exhaustion, surfaces SummaryRecomputeError (ERR_SUMMARY_RECOMPUTE) instead of a silent warn; seed/import recover it to a warning without re-writing.
• #3152 autonumbers are assigned after validation, so a batch that dies in validation consumes no sequence value (no number-range gaps).