Isolate and atomically publish farm artifacts per concurrent run

[SUaDtL]

Isolate and atomically publish farm artifacts per concurrent run

Severity: high | Confidence: 0.96 | Effort: L

Where:

• plugins/ca/tools/farm.ts:1751-1765
• plugins/ca/tools/farm.ts:1885-1891
• plugins/ca/tools/farm.ts:1900-1909
• plugins/ca/tools/farm.ts:1930-1966

Evidence: Each invocation mints runId but still truncates the shared ${reportDir}/farm-results.jsonl with writeFile(resultsStream, ""), appends all runs to that same path, and directly overwrites the same farm-report.json, farm-report.md, and diffs/<task>.patch names. There is no run-scoped directory, file lock, compare-and-swap, temp-file rename, or serialization across farm processes.

Impact: Two farm invocations using the same repository can erase each other's streaming receipts and final authoritative reports. A crash during any direct report write can also leave truncated JSON/Markdown, destroying restart evidence exactly when the report is needed for recovery and audit.

Recommendation: Make runId an ownership boundary: write streams, diffs, and final reports under a unique run directory; publish any stable latest pointer/index with a lock and same-directory temp-file plus fsync/replace. Never truncate a shared stream at startup. If shared aggregation is required, use a locked append protocol and retain immutable per-run sources.

Acceptance criteria:

• Two intentionally interleaved farm processes preserve complete independently parseable streams, diffs, JSON reports, and Markdown reports for both run IDs.
• Killing a farm process during each report-write boundary leaves either the prior complete artifact or the new complete artifact, never truncated JSON.
• Restart logic can enumerate and attribute every settled task to its run without relying on last-writer-wins files.
• Tests exercise concurrent start, concurrent task settlement, and concurrent finalization against the same repository.