[iamgp]

Why the change

The scan skipped most of a worksheet form — rows that list their answers as label/value pairs side by side, pairs whose label is merged across the columns it spans, single-row blocks, one-row tables of formulas, and tables whose data fills only the first column — so a lab worksheet yielded 8 fields where 36 are extractable, and accepting a table retired every suggestion below it.

Special things to note

• Recall over precision on purpose: the queue is ranked, so weaker extractions sit at the bottom instead of being dropped. The queue cap rises from 24 to 80 suggestions.
• Version number | 6 still comes back as a discriminator rather than a field, so the two never overlap.
• Table data rows are still not read as pairs, and a 2-row label/value block with no title is still a cover block (fields) rather than a table; the title is what makes it a table.

Change outline

The pair scan walks each row in both directions. Four guards decided what a pair is, and each one was dropping real form content.

 findKeyValueBlockSuggestions
-  valueCol = col + direction                    # a merged label lost its value
+  valueCol = past the label's own merge         # "label | (merged) | Yes"
-  if (a value follows) continue                 # a second pair hid the first
+  if (a value follows && it is not a pair) continue
+  if the value is the pair on the left          # "AB | AB | 08/02/2024"
-  if (size < 2) continue                        # single-row blocks dropped
+  if (size < 2 && not form-like) continue
+  isFormLabel(label)                            # sentences are labels next to their value
+  hasOwnValue(cell)                             # a merge repeats one value, not columns
 findCoverBlockSuggestions
-  valueCols.length < 2                          # one answer under a label row
+  valueCols.length < 1
 scanWorkbookForSuggestions
-  .slice(0, 24)                                 # the tail was truncated
+  .slice(0, MAX_SUGGESTIONS)

The table paths gained two measurements, and the pruning rule was wrong for open-ended fields:

 findTableSuggestions / findTitledTableSuggestions
   typedSampleCells  += values that start with "="     # a one-row table of formulas
+  measureFirstColumnBlock(startCol)                   # data may fill only column A
+  effectiveDepth = max(depth, firstColumnBlock.depth)
+  boundsEnd      = max(strictBlock.endRow, firstColumnBlock.endRow)

 dropSuggestionsCoveredBy
-  field contains the suggestion                 # "A23:F" runs to the bottom of the sheet,
+  field and suggestion start inside each other  # so containment retired every card below

The shapes that were being missed, from the reported worksheets:

row 3   Operator (Full name) | Alison Fernandes | Date | 07/02/2024
        -> 4 fields (was: date only, and operator dropped as "table data")

row 29  PBS aliquoted for Blank and NPC as per PRDSOP | (merged A:B) | Yes
        -> 1 field (was: nothing, the label's merge hid the value)

row 18  Incubator | 21LAB421 | 21LAB421        (merged B:D)
        -> 1 field (was: nothing, the merge read as extra columns)

row 66  Date last performed | 01/02/2023 | Date next due | 01/03/2024
        -> 2 fields (was: nothing, a single row never qualified)

Table 3: NTCs      Well# | CtΨ | CNΨ | CtRPPH1 | CNRPPH1  over one data row   -> table
Table 6 (A block)  MQCN NTC,Ψ | MQCN NTC,RPPH1  over formula cells             -> table
Table 1            Extract. Assay ID | IPC-EX Sample ID | ...  first column only -> table

 editor/src/
 └── lib/
     ├── suggestions.ts          # pair, merge, single-row, formula and first-column rules
     └── suggestions.test.ts     # 8 new cases