Continue OpenPresentation native compatibility testing on THIS Windows PC. Set this session's goal to the following:

Establish repeatable, safe, evidence-backed real Microsoft PowerPoint tests for current OPF exports: actual native rendering, editability, save/reopen, workbook edits, current-content reimport, font resolution/substitution and raster comparisons. Diagnose and fix reproducible export/harness defects in narrow reviewed PRs while preserving the user's Office session and all existing work. Distinguish editable DrawingML, measured geometry, browser paint, native visual fidelity and arbitrary round-tripping. Do not claim equivalence without the actual native evidence and stated tolerances.

First fetch OpenPresentation/opf main and read docs/handoff-2026-09-08.md IN FULL, docs/handoff-2026-09-10-wrap-up.md, docs/evidence/native-resume/README.md and docs/plans/text-placement.md. Fetch OpenPresentation/opf-pptx main and read docs/native-text-checks.md plus its native chart, metric and process harnesses. Review actual branches/PRs/checkpoints before acting; the wrap-up includes core #61, renderer #12, PPTX #16 and editor #11. Use GitHub commits as portable authority. Existing workspaces are under C:\Users\micha\Source\opf; inspect before reusing them and preserve dirty/untracked files. The four package repositories must be named siblings for coordinated scripts.

A SEPARATE agent on a Mac owns the larger ecosystem, product roadmap, release sequencing and public sites. You own Windows native harnesses, narrow native defect fixes and portable evidence. Use a distinct codex/windows-native-* branch, avoid modifying shared release/CI refs without coordination, and report exact source commits and focused PRs for the Mac owner to integrate. Do not publish packages or deploy sites as part of this testing task. The existing published set is core 0.9.0, CLI 0.7.0, renderer/PPTX 0.7.0 and editor 0.6.0; the new integration remains unpublished and requires coordinated sources/candidate tarballs.

Start by diagnosing Office readiness. The last warning from PowerPoint was: “To insert a chart, you must first close any open dialog boxes or cancel editing mode in Microsoft Excel.” The user had dismissed an earlier dialog and the COM Version probe worked, but a second ChartData.Activate() hung. A subsequent read-only COM/cleanup probe also hung. No Office process was terminated and fixture cleanup completion is unknown; an owned generated charts-saved.pptx may still be open. Do not assume an absent Excel process or null COM property means an empty/ready session.

Do not kill PowerPoint/Excel, call Application.Quit(), close all documents, discard unsaved work, alter global security settings, or loop through dialog-producing retries. Inspect the current state using supported capabilities, explain any exact user action that is required, and keep unrelated work moving. Close only a verified fixture presentation/workbook opened and owned by your run. Never close a pre-existing user workbook. If automation blocks, record it and stop that attempt.

The prepared PPTX chart harness now requires one explicit -EditSlide 1..8, uses indexed COM Count/Item collections, records progress, launches an owned hidden PowerShell helper and bounds it to 45 seconds (maximum 60). Its timeout kills ONLY that exact helper process, not Office or descendants, and never retries automatically. test/native-process-check.ps1 has real success, literal-argument, exit-17 and timeout checks. Read the scripts and docs for actual arguments; do not revive the old unbounded multi-chart loop. Make one controlled bounded chart attempt after the Office state is understood. Full edit coverage requires all eight slides in separate successful fresh runs for each supported Node runtime.

Recorded partial results: all eight original/native-saved charts opened, saved and reopened; 16 original/saved data imports and eight native PNG pairs agree. The first embedded-workbook edit returned, but activating the second hung. No final edited-deck comparison completed. This is partial evidence, not a full chart-edit pass.

The native accepted-text harness is prepared: Node 20.20.2 and 24.20.0 each generate 24 cases / 144 editable lines with identical package bytes and geometry. All four exact Carlito faces pass temporary Windows registration/removal. Full native text execution is still pending. Inspect and, if necessary, apply the same bounded-worker protections before running it. Temporary font installation must be removed in finally; preserve user/system fonts. Capture actual application/OS/build/font hashes, native output, renamed/edited current text, save/reopen results, reimports and independent ink masks. The documented text gate is nonzero pixel centers inside accepted bounds plus 0.1 reference pixel; do not widen it or clip/nudge consumers to hide failures.

Fresh metric fixture generation produced 48 Calibri reference decks but fresh native execution remains pending. Earlier counterexamples are still open: eight tab comparisons exceed 0.02 pt, maximum about 0.067383 pt; portrait/right Latency ink reaches x=497 beyond an accepted right edge of 496.8. Reproduce against exact source and fonts, keep failed artifacts, and fix shared causes instead of silently moving thresholds or changing source content.

Expand native checks to representative tables, quotes, code, metrics, scalar/rich text, charts, images, editing/undo where relevant and repeated export/reimport. Compare open substitutes with installed proprietary reference fonts without redistributing proprietary font bytes or outlines. Record reference hashes/environment, separate metric compatibility from visual and raster equivalence, and clearly label missing glyphs, substitutions, unsupported cases and OS/rasterizer differences. PowerPoint is a test environment only; the OPF runtime must remain open, local, provider-neutral and usable without paid services.

Preserve portable evidence in GitHub with raw failure/success logs, checksums, generated fixtures, verifiers, exact source commits and bounded scope. Skip Office ~$/lock files when collecting evidence. Run evidence-verification scripts against staged/Git blobs so manifests match what another computer can fetch. Keep the handoff current, include concrete reproduction commands and avoid relying on this PC's local paths in the portable record. Pushes, narrow PR updates and merges are authorized; coordinate shared product changes with the Mac owner. Report any unresolved native gate candidly.