본문으로 건너뛰기

ADR 0066 — Column-scoped data block

  • Status: accepted
  • Date: 2026-05-23
  • Spec target: XTL 0.1
  • Affects: language.md (Data Blocks section, new); evaluation.md (Render Phases); impl (parser, renderer); new error code

Context​

Up through 0.7.0 the engine treats the entire width of the data-row range as part of a single data block. renderer.ts captures every cell of the block's template rows via row.eachCell({ includeEmpty: true }) regardless of which column the cell sits in, and clones the captured cells into every expanded row. Rows below the data block shift down by the splice-insert count across all columns.

This row-wide assumption is not normative — spec/language.md defines "data block" only in row terms (the rows containing {{[col]}} expressions). The column scope of a block has been implementation- defined since 0.1.

The row-wide assumption is a latent bug factory for templates with content beside the data block on the same sheet — the "side summary table" pattern that's common in invoice / 거래명세서 / 정산서 workflows:

row 3:  헤더A  헤더B  헤더C  ...  헤더N  |  요약 헤더  값
row 4:  {{[a]}} {{[b]}} {{[c]}} ... {{[n]}} |  링키지랩 A  =SUMIF(...)
row 5:  합계   {{SUM([b])}}                |  링키지랩 B  =SUMIF(...)
row 6:                                     |  합계        =SUM(Q4:Q5)

Concrete failure modes observed at 0.7.0:

  • Issue #46 — silent data loss. OOXML shared-formula owner cells in P4–S4 (outside the apparent "data" columns A–N) are cloned into all 858 expanded rows verbatim, producing 858 duplicate owners of the same ref range. Excel sees this as corrupt and either drops cells or surfaces a repair dialog. Worked around in 0.7.x with unshareFormula (commit bcbe239), but the root cause is that outside cells were captured-and-cloned at all.
  • Issue #47 — formula ref staleness. Cells below the data block but in non-block columns (P5:S14 in the example) get row-wide shifted by spliceRowsPreservingMerges. The formula text in those cells (e.g., =SUM(Q12:Q13)) is preserved verbatim by ExcelJS, so after shift the row references no longer match the cells' new positions. xl3 makes no attempt to rewrite formula refs.
  • Replication noise. Static side cells (e.g., P4 = '링키지랩 A') get cloned into all expanded rows, producing N identical copies alongside the data — almost never the author's intent.

Two corrective design directions were considered: per-cell formula ref rewriting (matching Excel's "Insert Rows" behavior) or column- scoping the data block. Column-scoping is the simpler, more durable fix — it aligns with Excel's natural intuition (typing in column A doesn't affect column P) and removes the row shift for outside columns entirely, so the ref-staleness problem doesn't arise.

Considered Options​

A. Column-scope the data block (this ADR). The block's column range is the bracket-expression hull. Cells outside that range are preserved at their original row positions; the splice's row shift applies only to cells inside the block's column range. Single block per sheet (multi-block deferred to future ADR).

B. Row-wide block + post-splice formula ref rewrite. Keep the current row-wide capture. After splicing, walk shifted cells, parse formula text, adjust row digits in cell references. Matches Excel's Insert Rows semantics literally.

Pros of A: simpler impl; no formula parser needed; no risk of mis-adjusting refs (false positives from string literals, function names, etc.); outside cells never shift so #47 disappears by construction; aligns with the user mental model that the data block "belongs to" specific columns.

Pros of B: matches Excel UI behavior for "Insert Rows" 1:1; lets authors put any content in any column without worrying about scope.

Cons of A: behavior change for templates that put static text or formulas inside data-block rows in non-bracket columns expecting it to replicate — those cells now stay at the original row only (the user-helpful direction for almost all real-world cases).

Cons of B: formula parsing is fragile (cross-sheet refs, quoted sheet names, string literals containing cell-like patterns, named ranges); risk of incorrect ref adjustment is itself a silent-data- loss path; doesn't solve the "static text replicated 858 times" noise.

Option A chosen. Aligns with xl3's philosophy of small, predictable surface (ADR-0043, ADR-0048); avoids introducing a formula parser into xl3's scope; eliminates an entire class of latent bug (shared-formula corruption, ref staleness) by construction rather than by patching.

Decision​

The following becomes normative spec text in language.md's new Data Blocks section:

A data block on a sheet is the maximal rectangle [r_start..r_end] × [c_start..c_end] such that:

  • Every row in [r_start..r_end] contains at least one cell whose {{ ... }} body references at least one [Column] outside an aggregate function (a data-row cell).
  • c_start is the minimum column of any cell with a {{ ... }} expression body (data-row cell or aggregate cell — {{ COUNT() }}, {{ SUM([col]) }}, etc.) in the row range; c_end is the maximum. Aggregate expression cells in a data-block row are inside the block because authors place them alongside data cells to render per-row context (e.g., a customer name with a count total in the same row, replicated per customer).
  • The row range is the maximal run of consecutive rows satisfying the first condition.

Cells inside the rectangle are block cells. Cells outside the rectangle but on the same sheet are outside cells.

Block expansion. When the engine expands the data block to N records (after @filter / @sort / @top / @group / @repeat right apply), the block's row range expands from r_end - r_start + 1 template rows to N × (r_end - r_start + 1) output rows. The expansion affects only [c_start..c_end] columns:

  • Block cells are cloned into the expanded rows (one record per r_end - r_start + 1 rows).
  • Outside cells with row r < r_start remain at their original (r, c) position.
  • Outside cells with row r ≥ r_start are preserved at their original row r position; they do NOT shift downward with the expansion. Their formula references (if any) are preserved verbatim.
  • Cells inside the column range but in rows r > r_end (i.e., below the data block, in block columns) ARE shifted down by (N - 1) × (r_end - r_start + 1) to make room for the expansion.

A {{[Column]}} reference in an outside cell raises xl3/expression/bracket-outside-block at parse time (no implicit single-block fallback — outside cells must not reference data columns, since the cell does not iterate with the block).

For 0.7.x compatibility, the engine continues to assume a single data block per sheet; multi-block detection and the optional @block directive for explicit boundaries are deferred to a future ADR.

Add to evaluation.md "Render Phases":

Per-column shift on splice. The render phase that inserts rows for block expansion (renderDataRows) MUST apply the row shift to cells in [c_start..c_end] only. Cells in columns outside that range MUST appear at their original row positions in the output, regardless of expansion factor.

Consequences​

Behavioral change for templates with outside-column content. Templates that placed static text or formulas in non-bracket columns of the data-row range previously had that content replicated into every expanded row. After this ADR they appear at the original row only. This is user-helpful in all observed cases (issue #46 and similar replication-noise reports). No production user has reported intentional use of the row-wide cloning behavior.

Backward compatibility — narrow break. Templates without any outside-column content are unaffected. Templates with outside-column content render differently but the new rendering matches author intent. Migration notes added to CHANGELOG.md and the 0.7 → 0.8 section of docs/migration-0.x-to-1.0.md (when that guide is created per ROADMAP G19).

New error code. xl3/expression/bracket-outside-block added to the catalog. Raises at parse time when a [Column] reference (i.e., a data-row marker) appears in a cell that is not inside any data block's rectangle. G3's 30-day catalog-frozen clock (ROADMAP) resets along with the existing 0.7.0 additions; the engine's first new-code-free 30-day window starts from the latest of these additions.

Issue #46 resolution path. The unshareFormula workaround (src/renderer.ts, commit bcbe239) becomes redundant for the outside-column path. It is retained for defense against any future case where a shared-formula owner appears legitimately inside a data block's column range — a narrower scenario that was not the root cause of the reported failures.

Issue #47 resolution. Outside cells do not shift, so their formula references are not subject to row-number staleness. No formula parser or ref-rewriter needed.

Conformance corpus. Fixtures added:

  • 141-block-column-scoped-side-cells — side cells preserved at original positions across expansion
  • 142-block-column-scoped-side-formulas — formula cells with row refs in side area remain valid (#47 regression)
  • 143-block-shared-formula-side-cells — shared-formula owners in side area not duplicated (#46 regression)
  • 144-block-side-cells-after-block — distinction between in-block rows shifted (e.g., footer cells in block columns) and outside- block rows preserved (side rows)
  • 145-block-bracket-outside-error — [col] reference in outside cell raises xl3/expression/bracket-outside-block

These fixtures contribute to ROADMAP G1 (≥ 140 conformance fixtures); five new entries bring the corpus from 139 to 144 at 0.7.1 cut.

Multi-block deferral. Multi-block per sheet, the @block directive (with range-spec forms), and per-block directive scoping are explicitly deferred. They remain candidates for a future ADR once concrete multi-block use cases emerge. The current ADR sketches the column-scoping primitive that any future multi-block design will build on; that future ADR will need to address detection algorithm (strict explicit vs implicit cluster), directive scoping rules, and their interaction with @source / @repeat right.

References​

  • Issue #46 — duplicate shared-formula owners producing corrupt OOXML (silent data loss).
  • Issue #47 — formula references in shifted outside cells become stale.
  • Commit bcbe239 — unshareFormula helper introduced as a defensive fix for the inside-column case of #46; remains relevant for the narrow scenario where a shared-formula owner appears inside the block's column range.
  • ADR-0043 — Excel-native preference principle; this decision follows the "small predictable surface" preference over the larger "Excel-fidelity formula rewriter" alternative.
  • ADR-0046 — Cell formula preservation contract; this ADR is consistent (still no formula rewriting; outside cells stay where they are, so their refs need no rewriting either).
  • ROADMAP G1 — conformance corpus ≥ 140; this ADR's fixtures contribute.
  • ROADMAP G24 — silent data loss is a critical-bug category; #46 and #47 both fall under it. This ADR closes the structural cause.