feat(pptx): applyEdits — lossless surgical-edit API

[karthikmudunuri]

What

Implements applyEdits(source, plan, options?) — a lossless surgical-edit API that patches the original .pptx bytes instead of re-serializing a full deck. Everything not named by an edit comes out byte-identical to the source (masters, layouts, theme, embedded fonts, ppt/tags/*, notes, embeddings, and any untouched element), and the result opens in PowerPoint with no repair.

This removes the lossy round-trip that produced the custGeom / SVG-fallback / dangling-rel fidelity bugs and lets hosts drop their defensive cleanup. serializeDeck stays for the live editor and from-scratch decks; applyEdits is the lossless path for template-derived output.

API

const deck = await parsePptx(source);          // address elements by deck ids
const out: Uint8Array = await applyEdits(source, {
  title: "Q3 Results",
  slides: [
    { source: { slideIndex: 1 }, edits: [{ op: "setText", elementId: titleId, text: "Q3 Results" }] },
    { source: { slideIndex: 3 }, edits: [
        { op: "setChartData", elementId: chartId, categories: ["Jan","Feb","Mar"], series: [{ name: "Revenue", values: [10,20,30] }] },
        { op: "removeElement", elementId: sampleChartId },
    ] },
    { source: { slideIndex: 4 }, edits: [] },   // kept byte-identical
  ],
}, { onWarning: (w) => notifyHost(w.message) });

Elements are addressed by the same stable ids parsePptx returns; slides by 1-based template index. Call applyEdits in the same process as the parsePptx that produced the plan.

Ops supported

setText / clearText (preserve the template box + first-run styling, or rebuild from supplied runs), setChartData (in-place native-chart fill keeping type/colours and regenerating a consistent embedded .xlsx so Edit-Data still works), setTableData, setImage, removeElement, addChart, addDiagram, plus per-slide background and deck title.

How it works

• New elementLocationRegistry in the parser (getElementLocation) maps every element id → its verbatim source XML block, with no placeholder filtering (the existing elementSourceRegistry skips no-xfrm placeholders, which would break setText on titles).
• Output slides are built in plan order; reorder is just <p:sldIdLst> order. A repeated source slide is deep-cloned (slide XML + rels + editable deps) from a pristine view, so edits to copies never bleed.
• Slides the plan drops, plus any parts that become exclusive to them, are reclaimed by a package-wide reachability sweep; then reconcileDanglingRels + content-type pruning run as backstops.
• Unresolved element ids and unsupported layout-instantiation surface via onWarning instead of throwing.

Scope note: source: { layoutId } (layout-instantiation from scratch) is intentionally not in the lossless patch path — that's serializeDeck's job. applyEdits emits a layout-unresolved warning and continues rather than shipping a wrong slide.

Tests

New apply-edits.test.ts (11 tests) on a synthetic 4-slide template — no branded fixture needed — including the spec's acceptance test: edits text on 3 slides, fills one chart via setChartData, removes one sample chart, leaves slide 4 untouched, then asserts slide 4 + its media are byte-identical, the chart's embedded workbook reflects the new data, the package keeps its root rels / content types, and there are zero dangling rels.

Two real bugs were caught and fixed while writing tests: an rPr capture regex that stopped at a self-closing child element (corrupting run styling), and a GC step that deleted the mandatory _rels/.rels (invalid package, masked because tests only checked surviving rels) — now covered by an explicit assertion.

Verification

• pnpm typecheck ✅
• pnpm test ✅ (169 passed, 9 skipped — the skips are branded-fixture-gated)
• pnpm --filter @textcortex/slidewise build:lib ✅

Note: pnpm lint can't run in this workspace (@eslint/js isn't installed at the repo root — pre-existing), so static checking relied on tsc.

Minor changeset included.