[creedants]

Fixes the file-reading problems that five first-time testers hit in the second QA round on 2026-10-28. The two that gave wrong or misleading results come first.

• The summary's count of unnamed lines now counts every line and prints whole totals. Before, it could undercount and print a rounded total.
• A file whose separator puts a whole row into one period column is named, instead of quietly dropping rows. Two-column tables and periods that cannot hold a whole row are left out of that check.
• A text cell in a measure column is named by the file's own spreadsheet row, and only when the file proves that row. This holds after blank lines, CRLF, UTF-16 and Windows-1252 files, Excel sheets, and the app's query wrapper. When no row can be proved, the message names no row rather than a wrong one. The proof is a per-load stamp kept on the browser's DuckDB connection. Nothing new leaves the browser.
• Huge amounts and percentages are capped instead of printed as exponents.
• A month written without a day is read as the first of that month. Excel header cells are trimmed. German customer columns are left out of the suggested attributes. A file with too many columns is refused before the reader runs.

Checked with npm test (956 pass, 1 todo), npm run typecheck, npm run build, npm run build:cloudflare:free and npm run smoke:preview, and in a production preview with the testers' files and their variants. The million-row load time is unchanged from before the round.

Coordinated by claudeAgent/claude-opus-5-5. Written by opencode-go/deepseek-v4.1-flash. Reviewed by codex/gpt-6.1-sol.