dataforth(datasheet): Fix 2 — emulate QB single-precision rounding (26 -> 9 dirty)

formatMeasuredExact now applies Math.fround to the parsed value before toFixed.
The DOS QuickBASIC computed/stored these as single-precision floats, so the
last-digit rounding at the .5 boundary follows single, not double, precision.
Without it, double-precision toFixed flipped boundaries (9.9995 -> "9.999" vs
golden "10.000"; 46.85 -> "46.9" vs "46.8"; .45 -> "0.5" vs "0.4"; 3.3325 ->
"3.333" vs "3.332"). Verified each against the staged golden.

STAGE 3 re-validation: FINAL-TEST CLEAN models 68 -> 85 (+17), mismatches
26 -> 9, cert matches 2123 -> 2206. Zero regression — every remaining dirty
model was already dirty pre-fix; no previously-clean model flipped.

The 9 remaining are NOT rounding:
- 4 models (DSCA38-05/-1793/-19C/-19E): Supply Current retest data-vintage — the
  staged .TXT is an older test run than the DB latest-wins record; not a render
  bug, can't reconcile against an older sheet.
- 5 models (DSCA39-01/02/05/07/1950): STAGE 1 template artifact — the footer note
  "Standard output load for test is 250 ohms." was mis-captured as a truncated
  parameter row. A renderer-side fix was attempted but emitted the note for all
  current-output models (regressed 24 clean -C models), so it was reverted; needs
  a targeted STAGE 1 extractor fix instead.

Renamed dsca-clean68-models.json -> dsca-clean-models.json (now 85 models).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

projects/dataforth-dos/datasheet-pipeline/implementation/templates/datasheet-exact.js

@@ -327,8 +327,14 @@ function formatMeasuredExact(statusStr) {
     // char at index 4 is either a space (positive) or '-' (negative); start there
     // so negative signs survive (e.g. "PASS-4.2424060" -> "-4", not "4").
     const valueStr = statusStr.substring(4, statusStr.length - 1).trim();
-    const value = parseFloat(valueStr);
-    if (isNaN(value)) return { passFail, valStr: valueStr, value: NaN };
+    const parsed = parseFloat(valueStr);
+    if (isNaN(parsed)) return { passFail, valStr: valueStr, value: NaN };
+    // The DOS QuickBASIC stored/computed these as single-precision floats, so the
+    // value at the half-rounding boundary rounds the way single precision rounds,
+    // not double. Recover the single (Math.fround) before formatting — without it,
+    // double-precision toFixed flips last-digit boundaries (e.g. 9.9995 -> "9.999"
+    // here but "10.000" in the original; 46.85 -> "46.9" vs "46.8").
+    const value = Math.fround(parsed);
     const d = parseInt(decimalDigit, 10);
     const valStr = isNaN(d) ? value.toFixed(1) : value.toFixed(d);
     return { passFail, valStr, value };