Local investigations

A person proposes an investigation by typing / on a document line and choosing Investigate; its card sits at that place in the document. Someone selects Run on the card, which starts this run on their own local agent. A failed or stopped card's Retry reuses the same investigation id.

1. Read read_investigation before acting. Its sealed brief, source context and capability schema are authoritative for this run.
2. Use your normal repository instructions and configured tools. Keep the work within the authorized objective. State blockers rather than inventing access, results or a different source revision.
3. For performance comparisons, check correctness and measure under comparable conditions. Record units, samples, warm-up and aggregation. Distinguish observed, derived, simulated and qualitative data.
4. Return a short report, led by a # heading that names the finding (it becomes the title of the published child document), typed datasets, supported native view mappings and useful bounded text evidence. All required columns must be present in every row; nullable values are explicit and are not zero. Preserve source and capture provenance. Never include service credentials or indiscriminate transcripts.
5. Use tables, metric comparisons, bars or lines from the supplied schema. Do not generate frontend applications, executable chart configuration or live URLs.
6. Call submit_investigation_result. Repair validation failures and submit a valid candidate before finishing the turn. The connector handles durable publication. Do not claim a decision was accepted or start another investigation yourself.