fix(core): preserve cells when structurally editing ragged GFM tables

[changdapeng]

Summary / 摘要

Structural table edits (column drag-reorder, column delete, Delete-key range clear) dropped cells and misaligned the header on GFM tables whose rows omit the leading/trailing pipe; rows are now parsed GFM-consistently and touched rows are re-serialized in canonical form.

Motivation / 背景与动机

GFM allows table rows to omit their leading and/or trailing pipe, and remark renders such "ragged" tables as normal table widgets. But moveColumn / deleteColumn / the range clear parsed rows with split("|") + slice/filter, which assumes canonical pipes:

• on ragged rows the first/last cell was silently dropped;
• moveColumn additionally skipped rows whose parsed cell count fell short of the drag index — data rows were reordered while the header stayed put, corrupting the table.

Minimal reproduction (drag the 年龄/职位 grips to swap the two columns):

姓名 | 年龄 | 职位  |  |
| --- | --- | ---  | --- |
| 张三 | 25 | 提示词工程师  |  |
| 李四 | 30 | 测开工程师  |  |

• Issue: none filed — the defect and the expected deliverable ("真正修复表格列拖动问题的最小 PR，并配套少量稳定的 vitest/Playwright 回归用例") were described by the maintainer in the PR test(core): add table drag eval harness #102 review; the reproduction corpus from that PR is ported here as unit tests.
• Roadmap (docs/ROADMAP.md): n/a (bug fix)
• OpenSpec change: n/a (bug fix; the new module is internal and not exported from the package entry — no public API change)

Changes / 变更内容

• packages/core:
  • New internal module src/table-markdown.ts — pure GFM table-source transforms: splitRowCells (strips at most one pipe per edge, honors \| escapes via a character scan), serializeRowCells (canonical leading+trailing-pipe form; no cell-content substitution, so pre-existing zero-width cells round-trip byte-identically), moveTableColumn / deleteTableColumn / appendTableColumn / clearTableCellRange, and tableColumnCount. Two deliberate semantics, documented in the module: the delimiter row is identified by POSITION (always the block's second line, per the GFM spec) — a dash-only body row like | - | - | is data and must never be treated as a delimiter — and the column count is the max over all rows, matching the widget's own max-colCount rendering in toDOM() (cells beyond the delimiter count are preserved, not dropped). Every row is padded to that width before editing, which removes the silent row-skip; out-of-range/negative indices return the source unchanged; alignment markers move with their column. Pure functions, no DOM/state, unit-testable in isolation.
  • src/live-preview-table.ts — moveColumn / deleteColumn / addColumn / the Delete-key range clear delegate to those functions (addColumn's old regex append misplaced the new cell on ragged rows and was escape-blind); addRow sizes the new row by tableColumnCount instead of the header's AST cell count; separator checks go through isSeparatorLine (the regex is module-private now). The widget interaction state machine (editing locks, rAF polling, drag lifecycle, dispatch ordering) is untouched.
• Tests:
  • packages/core/test/table-markdown.test.ts — 39 cases: the test(core): add table drag eval harness #102 corpus ported as conservation assertions (row count, uniform column count, cell-multiset preservation, normalized pipes), alignment-marker follow, escaped pipes, out-of-range/negative-index no-op guards, byte-identical well-formed controls — including pre-existing zero-width cells (| a || c |), which round-trip untouched when a different column is edited — and dash-only body rows (| - |), which are treated as data positionally, never as a delimiter.
  • packages/core/test/live-preview-regressions.test.ts — 3 widget-level regressions driving the real grip-click + Delete event chain; assertions read editor.getDocument() only, no layout/coordinate dependency (jsdom reports zero layout, so coordinate-driven tests would be unreliable).

Testing / 测试

• pnpm test passes — 557/557 (was 515 on main + 42 new)
• Affected packages build (pnpm build for @floatboat/nexus-core)
• pnpm typecheck (pnpm -r exec tsc --noEmit) — zero errors; git diff --check clean
• Manual UI check in electron-demo (real browser layout, real mouse event chain): the reproduction table's 年龄/职位 drag swap, grip-select + Delete column removal, well-formed tables behaving byte-identically to before the fix, and adjacent-table isolation
• Before/after screenshots of the reproduction table drag: attached as a follow-up comment

Compliance / 合规自检

• CLA signed
• AI disclosure: the functional code in this PR is not primarily generated by AI. AI assistance, if any, is described below.
  AI-assisted notes / AI 使用说明: The functional code in this PR was primarily developed by AI (Claude Code) under my direction; I reviewed it and take responsibility for every line.
• New dependencies: none
• No build artifacts committed
• No secrets committed

Checklist / 自检清单

• Title follows Conventional Commits
• No public API change (internal module, package README untouched)
• Touched live-preview-table.ts → walked through the 12 Table Widget rules: the diff is confined to three method bodies' string transforms + the SEPARATOR_RE import; no event chain / editing lock / rAF / drag-lifecycle code touched (rules 1, 8, 10, 11, 12 domains all untouched)
• Not a new capability / breaking change → no OpenSpec proposal needed
• In scope per GOVERNANCE §4 (core bug fix)

Known limitation (out of scope)

The wide-table corpus case (viewport narrower than the table; the drop target can't be reached while dragging) is a layout-dependent UX issue, not a source-transform defect — it can't be covered by stable DOM-only tests and is left as a follow-up.

Relationship to the stale-offsets PR

Textually this PR and the stale-offsets PR edit the same structural-edit method bodies, so whichever lands second needs a mechanical rebase: the union is simply passing the other PR's live source (this.currentTableSource()) as the transform input, e.g. this.dispatch(deleteTableColumn(this.currentTableSource(), colIdx)). The fixes are logically independent.

[CLAassistant]

All committers have signed the CLA.

[changdapeng]

Before/after of the reproduction table (real browser, real mouse drag of the 年龄 grip onto 职位):

Before — ragged header (missing leading pipe) + trailing empty column, rendered as a 4-column widget:

After — columns swapped with every cell preserved; all rows normalized with leading/trailing pipes; the adjacent well-formed table byte-identical: