v0.9.0

Fixed

• An undetermined outcome no longer exits as a plain failure. A write whose response was lost carries both error and outcomeUnknown, and the harness deliberately judges unknown first when writing the audit row — the change may have taken effect, so a blind retry could apply it twice. The CLI guard judged error first, so the audit said "may have taken effect" while the exit status told a script it had not happened. The two layers now agree (exit 2, not 1), and a test pins the ordering so it cannot silently flip back.
• delete_artifacts reported an irreversible destruction that had not happened (bug class #13 — submitted is not completed). GitLab answers the project-wide bulk delete with 202 Accepted: it queues the work and later removes only the artifacts it considers eligible, keeping locked ones. The tool treated the call as done, returned priorState as "what was destroyed", and the audit row said ok. Measured on GitLab 19.2.1: 202 came back, and 100 seconds later all seven artifacts were still present — an audit trail asserting a deletion that never occurred, which is the worst shape this defect can take. The bulk path now returns outcomeUnknown, reports the figure as inventoryAtRequest (an upper bound, not a receipt), explains that the server decides eligibility, and the CLI exits 2. The per-job path is genuinely synchronous — verified removing real artifacts — and still reports a confirmed result.
• The per-job delete over-counted what it destroyed. A job's log is listed among its artifacts but is not removed by that endpoint: after a real delete, every job.log (file_type: "trace") was still there while priorState claimed all seven files were gone. Traces are now excluded from the destroyed count, and fileType is carried on every artifact row so a caller can tell them apart.
• The dry-run of that HIGH-risk delete could never say how many artifacts it would remove. It read currentCount from the inventory's total key — a key the listing envelope deliberately does not have — so the one number an operator needs before an irreversible deletion was always null. It now reads artifactsFound, which counts every artifact found rather than just the returned page.
• Artifact sizes are integers (bug class #2). sizeBytes and totalBytes rendered as floats (374.0) although GitLab reports ints; the storage RCA's byte totals had the same defect from 0.0 accumulator seeds. An earlier round fixed only the write path and the changelog implied the whole surface — the read path was still wrong.
• The CLI exits 2 on an undetermined outcome, matching the rest of the line: emit_governed only knew about {"error": ...}, so a queued-but-unconfirmed write exited 0.