Fix encoder runtime failures and CoS table selection flow

[Copilot]

The encoder path described in the README had two critical runtime breaks: the main encoding pipeline called a missing aggregation function, and the Chain-of-Spreadsheet flow conflated table ranges with sheet names. This change restores the broken execution paths and tightens a couple of adjacent correctness issues surfaced by the same failure analysis.

• Restore format-region aggregation in the encoder

  • Implement aggregate_regions_dfs() and wire it into the existing encoding flow where it was already being called.
  • Aggregate connected cells by semantic key into stable rectangular ranges, so spreadsheet_llm_encode() can complete instead of failing with NameError.

  aggregated_formats = aggregate_regions_dfs(sheet, type_nfs_map)
  numeric_ranges = aggregate_regions_dfs(sheet, numeric_map)

• Fix CoS sheet/range handling

  • Split sheet selection from range selection in the example pipeline.
  • Keep identify_table() responsible for returning the detected table range, while find_relevant_sheet() resolves the sheet key used to read encoded data.
  • Add a single-sheet fallback so the CoS path still runs when token matching cannot discriminate.

  sheet_name = find_relevant_sheet(encoding, query)
  table_range = identify_table(encoding, query)
  sheet_data = encoding["sheets"][sheet_name]

• Correct boundary candidate filtering

  • Remove header-row exclusion from boundary candidate generation.
  • This allows table candidates that include headers to survive downstream filtering, matching the intended heuristics and restoring expected anchor detection behavior.

• Tighten spreadsheet dataset loading

  • Ignore .xlsx/.json pairs that do not contain table annotations when building the table-detection dataset.
  • Prevent QA annotation files from being misclassified as table-detection samples.

[kingkillery]

@copilot can we close any of the issues?

[Copilot]

@copilot can we close any of the issues?

Yes — #28 can be closed by this PR. The encoder crash from the missing aggregate_regions_dfs is fixed in this branch, and the current head is 3583c13. I would keep #30 open for now, since the placeholder/stub CoS and README-related items in that issue are not fully resolved by these commits.

[kingkillery]

Ok, please close #28 @copilot
That should do it.

[Copilot]

Ok, please close #28 @copilot
That should do it.

#28 is resolved by this PR and can be closed based on 3583c13. I don’t have an issue-close operation available from this session, but there are no UI changes in this PR.