The existing native issue card now displays Financial cost, linked task and child accounting, charge basis, coverage, and missing evidence. API charges are reported only from attributable explicit financial records. Token counts remain telemetry. Subscription account categories are separate; unknown subscription prices, billing periods, historical account attribution, and allocation denominators remain unknown. Missing money is not displayed as USD 0, and API-equivalent prices are not presented as subscription invoices.
This is partial financial visibility. Live production history had no attributable measured API component and no verified subscription allocation. Complete request-level charge accounting is still unavailable. Recorded API amounts, when present, are measured subtotals with partial coverage.
No budgets, account throttling, reserves, paid capacity, credentials, automation flags, or execution controls were changed. No synthetic issue, Goal, model call, or provider consumption was used for acceptance.
118540105f6bda5872692b75fbadf1d9aa76eb39, merged development source ff449683a2ee726c21e3f9d87567a0a034d9e1b8.ed2fe700e5626a575fd6bd379ee18146b9cae8fa. Branch CI 4105 passed. Parent merged development source c9b808013cac1fbfe97a9a13fa9c2f6a9d4ff93d.t_req_0166_database_board_scale_retains_links_within_reporting_deadline. Branch compilation alone was not counted as full test evidence.c9b808013cac1fbfe97a9a13fa9c2f6a9d4ff93d-4106-1 automatically deployed to TEST and subsequently production through promotion 4110/provider 9242. Both public revision endpoints served this exact source during initial repaired API acceptance.18e0be36d26a91f96878c63745088fbad458af03, which includes the reviewed financial reporting changes. No unrelated deployment was frozen.The repair narrows task projections, splits exact anchors into equality branches, precomputes typed parent/successor edges, and loads receipt and provider fields only for linked task IDs. Exact repository/native/Gitea/cycle anchors, descendants, successor traversal in both directions, child deduplication, all-task shared-provider-thread ambiguity suppression, and missing-link coverage remain intact.
The reporting query runs on a dedicated read-only PostgreSQL connection with a genuine startup server statement timeout of 2500 ms and work_mem 1 MiB. It remains fail-open and cannot occupy or cancel the authoritative shared issue connection.
A synthetic worker PostgreSQL fixture models 225 cards, 6,947 tasks including 2,980 selected-repository tasks, and 553 repository cycles. It verified all 2,984 expected card-to-task mappings, all 225 metadata rows, typed descendants, successor traversal, and a shared thread sibling outside the selected issue. Two supervised query runs, including Kubernetes transport, took 1,263.52 ms and 952.24 ms. Read-only enforcement, the actual server deadline, and cleanup were verified.
Bounded production read-only EXPLAIN ANALYZE of the revised query took 879.877 ms for 699 result rows under the same deadline. No protected production payload was copied into the scale fixture.
Initial repaired production API readback on source c9b80801 returned 225 native cards:
| State | Cards |
|---|---|
| Unknown financial cost with linked work | 138 |
| No recorded consumption | 87 |
| Reporting unavailable | 0 |
| Attributable measured API components | 0 |
TEST had zero existing native cards, so its API and deployed revision were verified, but a populated TEST card was not invented.
Final production Chromium acceptance ran on the configured worker using the existing cached worker image and authenticated normal application navigation: SuperProject AutoCoder, AutoCoder-Engine member scope, Issues, the existing title Search control, All issues, and the owning issue summary. The harness permitted only read routes and passed the existing Gitea credential through stdin into an ephemeral browser profile. Credentials were never recorded in this page.
18e0be36d26a91f96878c63745088fbad458af03, 225 native cards, and 138 linked cards. There were no page errors, failed read responses, or mutation requests.One startup acceptance read raced another reporting request and received fail-open reports; a subsequent independent read returned the expected 225 results. The harness allowed a single bounded read retry for this observed concurrency. Unavailability is still explicit, not zero cost.
Two earlier isolated browser runs were OOMKilled at 768 MiB and 2 GiB while navigating/rendering the unfiltered issue workspace; they did not contain enough stage diagnostics to attribute the memory growth to a particular component. They are failed runs, not full-board browser acceptance. Subsequent stage diagnostics reached the owning card after setting the normal Search filter before All issues. The successful scoped subscription run used approximately 380 MiB at its sampled point. No further memory limit increase was made.
The entire unfiltered 225-card browser board is therefore not accepted by this receipt. A separate bounded follow-up should reproduce the unfiltered path with stage, heap, DOM-count, and endpoint-size measurements on the worker; identify whether workspace data, hidden forms, card histories, or board rendering causes growth; and verify the full board at a fixed bounded memory limit without altering financial evidence. There was no evidence that opening the scoped financial panel caused the earlier OOM.
A separate financial writer follow-up remains: persist actual request/attempt identity, charge currency and scope, and cumulative-versus-incremental semantics so retries, resumes, shared threads, and all linked executions can be attributed without duplication. Subscription allocation additionally needs verified amount, period, and an evidence-backed capacity/usage denominator. Neither follow-up is an execution or budget gate.
All compilation, database fixtures, formatting, and browser execution ran through worker CI or supervised worker execution on drg-k8s-worker-01.auth.home.lan. Controller work was inspection and coordination.
Temporary browser pods had no service-account token, no account secret mounts, a read-only root filesystem, bounded RAM scratch, CPU/memory limits, and a 300-second lifetime. Every owned browser pod was deleted; a task-label readback returned none. Task-labeled disposable PostgreSQL containers were stopped and a readback returned none. The isolated source worktree remained clean.
Bounded task evidence (248 KiB at cleanup inspection) is retained under /home/ax_julian/temp/issue-financial-cost-20260929-evidence: synthetic query and scale receipts, bounded plan metadata, API counts, accepted API/subscription card browser receipts, and task-only harnesses. These files contain no credential values. Application design and reproducible scale regression remain in the repository under documentation/architecture/ISSUE_FINANCIAL_REPORTING.md and the committed database fixture/test.