bug: dashboard cost widgets round sub-cent totals to $0.00

Finding: In a self-hosted Langfuse v4.54.0 deployment, an external reporter says dashboard cost widgets show $0.00 for non-zero totals below about half a cent, even though the stored cost and the trace detail show the correct value. This is an external report; KFA did not reproduce it. Source: https://github.com/langfuse/langfuse/issues/18459 Reported symptom and environment: The reporter runs a cheap model (about $0.0006 per call). On the Home dashboard the Model costs card shows Total cost $0.00 and every model row also shows $0.00. The trace detail for the same call shows $0.000601, and the reporter states the ClickHouse total_cost value is 0.000601, so the aggregate is stored correctly and only the display drops it. Reported versions: self-hosted v4.54.0, Python SDK langfuse 4.17.0 with cost sent via cost_details; the reporter says the bug is in the web UI and not SDK-specific, and that main as of 2026-10-08 has the same costFormatter. Reported cause (reporter hypothesis, not maintainer-confirmed): The reporter attributes the behaviour to a change in #15866 after which costFormatter in web/src/utils/numbers.ts always rounds to two decimals. Call sites named as affected are ModelCostTable.tsx (total and per-model USD), ModelUsageChart.tsx (totalMetric) and UserChart.tsx (totalMetric). The reporter distinguishes this from #15109 (session rollup returns 0) and #16077 (root SPAN carries no cost). Reproduction as reported: ingest one generation with cost_details total 0.000601 (or use token usage on a model priced at a fraction of a cent per call), open the Home dashboard, and compare with the trace view. Proposed mitigation (reporter-sourced, unverified by KFA): keep two decimals for costs of $0.01 and above, and for nonzero costs below one cent use roughly three significant digits, capped at 10 decimals. The reporter says existing #15866 test values stay unchanged and lists new expected cases such as 0.000601 to $0.000601 and 0.004 to $0.004. The reporter says they run this patch on their own self-hosted instance and offers to open a PR. Status: No maintainer response, merged fix or release was captured in the supplied material. Closure or a PR would not by itself prove a shipped fix. Until maintainers respond, treat the cause and the patch as unconfirmed. If you see $0.00 on dashboard cost widgets, compare against the trace detail value before assuming data loss. Limitations: single external report, partial comment coverage, no independent reproduction, and no KFA reproduction.