[tnunamak]

What

The run detail page now shows records a run could not store or account for:

• the run's rejected-record count and the rejected-record receipts for the run (stream, reason code, first input index, payload size);
• the unaccounted count per STREAM_EVIDENCE stream.

Why

A run can succeed while the store permanently rejected some records (per-record isolation keeps the cursor moving), or while a stream left keys unaccounted (spec behavior). Both are recorded server-side, but the run page rendered neither, so such a run looked clean.

This PR keeps both behaviors and makes them visible. It does not hold the cursor and adds no table.

How

Data already on main:

• run.batch_ingested carries records_permanently_rejected, and the terminal event carries the run total.
• run.stream_evidence_declared carries outcomes.unaccounted.
• GET /_ref/connections/:id/record-rejections lists receipts, oldest first, with no run filter.

One timeline page holds only part of a long run, and the terminal event is emitted last. A count summed from one page can therefore miss the batch and terminal events, and a page without them is not evidence of zero. So the run total comes from the run status, the same window-independent source as known_gaps:

• Server (one small addition): GET /_ref/runs/:runId passes the terminal event's records_permanently_rejected through, only when it is a non-negative integer. It is absent while the run is active and for runs that predate the field.
• Console: summarizeRunRejections returns one of four states and never reports a zero it cannot prove. A terminal event of any kind, including run.abandoned, can supply the total.
  • final: the run status total, or the terminal event's total. A reported zero renders no section.
  • so_far: the run is still active and the whole timeline is loaded, so the batch sum is the total to now. Liveness is the page's own (run status, else the timeline envelope's terminal status); the summary keeps no list of terminal events for it. A finished run with no total, such as an abandoned run, is never so_far.
  • at_least: part of the run is not on this page; the batch sum is a lower bound.
  • unknown: nothing readable. The section says the count is not available here.
• Unaccounted streams come from evidence events, which have no window-independent source. When the page is not the whole timeline, the section says streams on other pages are not listed. Several evidence events for one stream are merged.
• Receipts: read when the count is positive or unknown (an unreadable count is when they are the only evidence), and the connection is known; a known zero reads none. The pager follows the cursor up to 10 pages. If the cap stops it with pages left, the section says the list is partial. If receipts cannot be read (fetch failure, or no connection id), it says the list is unavailable.
• The earlier "The other records were stored." is removed: retryable and unaccounted records are not stored either.

Mixed versions: a new console with an older server has no status total, so it shows so_far, at_least or unknown from the events it has, never a silent zero. An older console with a newer server ignores the extra field.

Tests

• run-rejections.test.ts (18): each state, a reported zero, the terminal total beating the batch sum, an abandoned run with and without a total, a finished run whose terminal event is not on the page (never so_far), events without the field, evidence merged by stream, receipts narrowed to the run with the cursor kept, and the page's receipt read through loadRunRejectionView (old-server status with a partial timeline reads the receipts; a known zero, or no connection, reads none; a failed read gives no list).
• rejected-records-section.test.ts (11): count and receipts, unaccounted count, no "other records" claim, lower bound, in progress, unavailable count, an unknown count that still lists receipts, other-page streams not listed, receipts unavailable, receipts partial, nothing rendered for a known zero.
• run-status-body-builders-pure.test.ts and ref-run-status-route.test.ts: the total passes through from the terminal event (also a reported zero), and is omitted when absent or invalid. The route test uses the file-backed database; PostgreSQL was not run, since the new field reads terminal.data, the object known_gaps already comes from.

The new console tests fail on main, where the modules do not exist.

cd apps/console
npm run test:files -- "src/app/(console)/lib/run-rejections.test.ts" "src/app/(console)/syncs/[[]runId]/rejected-records-section.test.ts"   # 29 pass, 0 fail
npx tsc --noEmit                                                                                                                              # no errors in the touched files
cd reference-implementation
node --import tsx --test test/run-status-body-builders-pure.test.ts test/ref-run-status-route.test.ts                                           # 22 pass, 0 fail
npx tsc --noEmit                                                                                                                                # exit 0

ultracite check reports no lint-rule diagnostics for the touched files. It reports formatting errors (it expects tab indentation, while existing files in these directories, including unmodified ones, use spaces) and unused-suppression warnings, all present on main.

Limits

• Receipts carry stream, reason, index and size, not record keys. A receipt stands for a distinct rejected payload, so the receipt count can differ from the rejected-record count.
• The receipt pager stops after 10 pages and says so.
• Other counts on this page are also summed from the loaded timeline page and are not changed here: checkpoint staged/advanced/commit-failed, skipped streams, interactions, {events.length} events, the terminal known gaps, and the failure and stderr sections.

Assisted-by: AI