Claude Code transcript

Everything is dispatched and moving. Here's where things stand.

Your AppleScript question: yes, confirmed. scripts/app_export_pdf.sh does it standalone, and scripts/visual_diff.py has the same mechanism built in — it opens a renamed copy of the fixture in the real app via AppleScript, exports a PDF, closes without saving, rasterizes at 150dpi, renders our viewer via Playwright, and writes side-by-side composites. I verified it end-to-end today on fixtures/templates/keynote/20_BasicBlack.key, and all three subagents are using it as their fix-verify loop (on separate viewer ports so they don't collide).

The three fidelity agents are running, each seeded with yesterday's ranked campaign findings plus your new template fixtures: - K (Keynote + shared drawables): master/layout placeholder inheritance (decks rendering blank/unstyled), master background fills, PDF-backed vector images. - N (Numbers + shared tables): number formats, stored column widths, alignment — plus your Invoice/Calendar templates. - P (Pages + shared text): page-layout docs rendering empty, the G5 acid checklist sweep, and your pagination/margins miss (I relayed it as top priority; good news — pageSize/pageMargins/orientation are already in the model and emitted by the converter, so it's pure viewer rendering work).

Model stewardship per your instruction: all three agents now route model/schema changes through me. First proposal already came in from N (thousands-separator flag, proto-sourced from TSK.FormatStructArchive.show_thousands_separator) — I approved and landed it as CellFormat.grouping (commit 9cd782c, TS + Rust in sync, typecheck and build green). One flag for you: commit 2ed592d ("compact flat points arrays for curve elements", marked Peter-approved) landed outside my approval loop — if that was you in another session, all good; if not, tell me and I'll review it.

Corpus-wide census (I ran the converter over all 964 modern fixtures since the crawl files have no preview.pdf for text comparison): text extraction is broadly healthy, but there are 1,167 table-degraded warnings across 218 docs in four concrete decoder classes — missing v3 cell-storage decoder, pre-BNC tiles, undecoded uid-based merge ranges (one doc has 996 covered cells), and v4 rich-text cells being dropped. Ranked list with worst offenders is dispatched to N.

Viewer shell (my lane, committed): the whole client area now accepts drops at all times — full-viewport "Drop to open" overlay, drop replaces the open doc, and parsing no longer blanks the screen (toast instead); plus dark-mode chrome with paper-white document areas, sticky header, landing format chips, and a JSON download button. Gate was 6/6 after my commits; the currently failing pages gate test is P's in-flight pagination DOM change and they've been told to own it.

Also committed your G2/G5 goldens + acid checklist (following the G1 precedent); I left the Apple template binaries untracked deliberately — they're Apple-authored content, and fixtures/templates/provenance.json is committed instead. I'll consolidate each agent's report (including their hand-built fixture wishlists for N/K) as they finish.

K's report is in — three committed wins, plus a cross-cutting style-resolution defect that lands in P's lane. As model steward I'll make the design ruling (fix belongs converter-side: the model contract says emitted styles are resolved, so absent-key = app default only holds when the chain was fully walked), then route it.

K checked in with three committed fidelity wins: canvas text was rendering 33% oversized everywhere (pt-vs-px unit mismatch, fixed with a 0.75 scale on the text layer), shape text sat a full shape-height below its geometry, master-slide "ghost" placeholder prompts leaked behind real titles (swapped args in an overlap check), and PDF/vector background fills now paint their real gradients. The RIPE test deck's slides are reportedly near-pixel-perfect now.

K also found a cross-cutting bug that explains a lot of tiny-text decks: pooled styles that only carry one override (say, font color) lose their inherited font size at emission, so the viewer falls back to browser defaults. Per the model's design contract, emitted styles must be fully resolved — so I ruled the fix converter-side (the style resolution walk is missing the theme/master hop for placeholder-owned text) and routed it to P, who owns that pipeline, with K on call for the Keynote-side archive specifics. Both agents are continuing on their remaining backlogs.

The tree has a three-way mismatch on curve points (TS model + viewer expect flat arrays; Rust still emits [{x,y}]), breaking every explicit-path shape and making visual_diff unreliable for everyone. As steward I'll land the Rust side atomically — model.rs struct plus mechanical emitter flattening in one commit so the tree never sits broken. First, the ratified contract and current Rust shape:

The model design review is done and committed as docs/model-review.md, with the urgent unblock landed first. Summary:

Unblock: K flagged that the tree had a three-way mismatch on curve geometry — the TS model and viewer had moved to compact flat point arrays, but the Rust serializer still emitted [{x,y}] objects, so every explicit-path shape rendered as an invalid path. I landed the Rust side atomically (enum → Vec<f64>, emitters flattened, normalize_path on coordinate pairs), verified flat output on a real fixture, rebuilt the wasm, gate 6/6. K has already refined the degenerate-axis handling on top — the shared-file coordination is working as intended.

Design review, against your three criteria:

• Compactness — strong, with two leaks. The measured decisions (pooled styles with omit-default, dense row-major grid with bare-scalar cells, media bytes out of band) are right and I recommend no churn there. The leaks: warnings emit per-instance (one corpus doc carries 376 near-identical table-degraded rows — I'm queuing converter-side aggregation with a count field), and TableModel.rows/columns are positionally dense, which is 28k mostly-empty objects on a big sheet — I ratified truncated/omitted-when-default arrays and told N to emit accordingly.

• Library usability — good bones. A consumer can render or extract text without any style lookups, and the untagged-union costs are documented and bounded. A few type-hygiene nits (bare string where HexColor/Fill/Rect belong) are noted for opportunistic fixes.

• Inheritance — this is where the real work was, and your instinct was right. I ratified one rule: the viewer never walks an inheritance chain. Three violations found, all now assigned: (1) pooled styles weren't resolved through the theme/master hop before omit-default comparison — that's the tiny-text bug K found, converter-side fix in flight with P; (2) Keynote slides said "background absent = go look at the master" and made the viewer invent master-compositing rules (the ghost-prompt bug was exactly that) — I landed Slide.masterDrawables (a resolved, pre-filtered underlay the viewer paints verbatim) and made background always-resolved; K is wiring extraction, which should also fix Home.key's blank slides; (3) same treatment for Pages layout canvases via FloatingPage.templateDrawables, assigned to P. The one place the viewer legitimately composites is word-processing page furniture — pages don't exist until viewer-side pagination — and that's now documented as a bounded name-lookup, not a chain walk. Table role-default styling was reviewed and deliberately kept as-is.

All model changes landed additively with the tree green (cargo + strict tsc + gate all passing), and each agent got its slice of the dispatch list.

No harm done — verified just now, the whole tree is healthy: release build clean, all 14 Rust tests pass (including the G1/G2 golden-output tests), the strict model typecheck passes, and the viewer gate is 6/6.

Your other agent's work was actually good — the flat curve-points contract it put into primitives.ts is a real compactness win and I ratified it. The only trouble it caused was landing one-sided: the TS model and viewer moved to flat pairs while the Rust serializer still emitted {x,y} objects, so for a window every explicit-path shape rendered as a NaN path. I closed that by landing the Rust half, and someone (K or your agent, per commit 3dc30b6) then finished the viewer decoder and wired tsc --noEmit into the gate — so this exact class of cross-language contract drift now fails npm test instead of shipping. The one other side effect, a stale test literal my masterDrawables field broke, was caught and fixed too (b9a2c73).

The brief test failures I saw mid-check were just races against P's live edits — on the settled tree everything passes. Meanwhile the agents have been landing real work: P shipped the text-inheritance fix (runs now inherit paragraph char properties — the tiny-text bug), sections/page-master headers/footers with live page-number fields in the paginated viewer, and N shipped the v4 rich-text trailing-slot fix, merged-region decoding from all three sources, and the ratified truncated rows/columns emission. K is on the masterDrawables extraction. I'll keep consolidating as their reports come in.

P's final report is in — big wins (page-layout empty fix, pagination+margins, the cross-cutting style-inheritance fix, G5 sweep) with clean gates. Before consolidating: one risk to check — P changed shared text.ts to emit px-for-pt, but K had earlier added a 0.75 scale in drawables.ts for the same root cause. If both survive, canvas text is now 25% too small. Verifying:

Round 1 is complete for two of three agents, round 2 is running, and my steward queue is clear. Consolidated status:

Pages/text agent — finished, all gates green (conformance sweep over all 1,248 fixtures: no defects). Headlines: page-layout docs no longer render empty (a one-line deprecated_storage fallback unlocked template text); word-processing docs now paginate into real page frames with margins honored — your priority item — with the KVKK doc matching Apple wrap-for-wrap; the cross-cutting tiny-text defect is fixed at the converter (runs merge over the paragraph's char-properties chain, so the RIPE deck's pooled styles all carry real font sizes now); and a long G5 acid-checklist sweep landed (headers/footers, live "Page 2 of 5" fields, list restart, custom checklist bullets, tabs, absolute first-line indents, footnote marks).

Keynote/drawables agent — finished, same gates. Home.key went from completely blank to all four slides matching; the RIPE deck is near-pixel-perfect on its tested slides. It extracted the masterDrawables contract (compositing now converter-side, keynote.ts just paints), added image-mask cropping, PDF gradient sniffing for vector backgrounds, and corrected two documented format facts with fixture proof (geometry angle is degrees, gradient angle is radians — both provenance-tagged). The brief double-scaling overlap with P's px-units change was caught and reconciled by the agents themselves.

Your fixture offer — both agents left build checklists: K wrote a complete G6 Keynote acid deck checklist (11 slides, one feature each) in .scratch/crossval-campaign/keynote/summary.md, and P asks for a small WP doc with {two-column section, drop cap, right-paren lettered list, footnote near a page boundary}. Building those would directly target the remaining gaps.

Steward work since: P's three model proposals ruled and landed (numberSurround for "(a)"-style lists, PagesSection.columns, footnotePlacement); and the review's Leak A is fixed — warning floods now aggregate at emission (the 376-row worst-case doc emits 2 rows with counts), with the viewer panel showing true totals and example paths.

Running now: round-2 Pages agent (extraction for the three new fields, the residual Keynote font-size resolution, 00C full-bleed cover) and round-2 Keynote agent (G2 preset fills — the pentagon/oval/parallelogram P spotted — tick rotation origin, vector-art preview fallback), plus the original Numbers agent still working through the table decoder classes. I'll consolidate as they report.

For your fixture-building offer, two ready checklists now exist: - G6 Keynote acid deck (11 slides, one feature each — placeholders, bullet levels, dark master, masked/rotated photos, shape sampler, valign/insets, slide-number pair, table+chart, vector image, skipped slide, notes+build): in .scratch/crossval-campaign/keynote/summary.md. - G6 Numbers acid sheet (formats table incl. control cells like checkbox/star/slider, exact widths/heights + hidden row/col, merge sampler, style table with per-edge borders and banding, table-name visibility pair): in the Numbers agent's report — I can write it into fixtures/golden/G6-numbers-checklist.md if you want it in the repo to build from. - Pages needs only one small addition: a footnote near a page boundary (everything else on the wishlist turned out to already be in G5, drop cap included).

Still running: the two round-2 agents (Pages: new-field extraction + drop-cap proposal + Keynote font-size residue; Keynote: G2 preset fills + golden re-sync, tick rotation, vector fallback, now charts). I'll consolidate when they report.

The drop-cap model field is landed and the Pages agent is extracting and rendering it against G5's big "T" now — the proposal was proto-verified end to end (storage keys paragraphs to DropCapStyleArchive via attribute-table field 28; G5's active style is a text cap, 3 lines, 1 character, 0.8 glyph scale, HelveticaNeue-Medium).

One correction to what I told you earlier, courtesy of the agent's archive scan: G5 does not actually contain a two-column section — the saved file has a single SectionArchive and no column styles (Apple's own export shows none either), so my "columns are already in G5" claim was wrong. The agent verified the new SectionColumns extraction against a crawl-corpus document with a real two-column body instead. So if you do build a small extra Pages fixture, the two worthwhile items are: a two-column section region and a footnote near a page boundary.

Two small diffs — a paragraph gained a resolved pStyle and the para pool grew by one — which is exactly the signature of P2's in-flight style-resolution work on Pages. Their brief binds them to verify-then-re-sync that golden before their final commit, so I'm leaving it to them rather than racing their edits.

Status: Keynote round 2 is complete, all its gates green. The headline is a genuine cross-app converter fix: drawable styles never walked the TSS.StyleArchive parent chain, so any shape wearing its theme's default look emitted an empty style — that's why G2's pentagon/oval/parallelogram were unfilled. The chain walk landed with preset disable-marker semantics (empty-pattern = no stroke, disabled shadow = none), plus a second real find: TSD group children are stored group-local, so our §3.4 re-basing was double-subtracting and flinging grouped drawables to the canvas corner. Also fixed: the Briefing tick marks (an inline-baseline CSS issue, not rotation), placed vector art now rides preview twins with a neutral placeholder fallback, ppd title sizes match after P's styles work, and line/area/scatter charts draw actual polylines with markers instead of bars.

Now running: Pages round 2 (drop-cap extraction/rendering, new-field extraction — footnotes at page bottom already found and wired to the model, per its interim report) and a fresh Numbers round 2 taking per-series chart colors (so the Running Log pace line goes Apple-orange), the v3 cell-storage decoder (the 376-warning census class), residual pre-BNC docs, and the Calendar row-height mystery. I'll consolidate when they land.

While that sweep runs (16 decks sampled across the size distribution, each rendered against its embedded Apple preview), a note on what you're seeing: the round-1 campaign only verified a handful of decks — the corpus-wide Keynote reality is exactly what you describe, and now we'll have composites naming the dominant failure classes instead of "most are disastrous". When the sweep lands I'll triage the composites by eye, rank the failure classes, and spawn Keynote round 3 aimed at the biggest ones.

Two practical notes meanwhile: make sure you hard-refresh (or restart the serve) — the dist was just rebuilt with today's full stack of fixes, including the group-local re-basing and theme-default fills that would have made many decks look very wrong this morning. And if a specific deck offends you most, drop its filename my way — a named disaster beats a sampled one.

Keynote round 3 is off and running against the triaged list, with the full 16-deck sweep now complete as its evidence base (each deck has an Apple-preview-vs-ours composite it can regression-check against).

The plan of attack it's carrying, in order: truncated text boxes first (the fix is semantic, not cosmetic — Keynote auto-grows plain text boxes and auto-shrinks placeholder text on overflow, while we clip at the stored height with taller browser font metrics; it'll pull the real auto-fit flags from the protos), then the LISA-style geometry blowups (diagnosed from the JSON before touching code, since that one smells like a slide-size-era issue rather than styling), TIFF transcoding in the wasm media path so masked TIFFs stop rendering as broken placeholders (with license-recorded crates per your repo rules), the missing master gradient, the ghost subtitle on the BICEP3 deck, and a standing watch for centering drift across every composite it produces.

Meanwhile Pages round 2 and Numbers round 2 are still working their lists. Your drag-testing is genuinely useful as a second validation channel — if you hit a deck that still looks wrong after this round lands, its filename is the most valuable thing you can hand me.

Two more model rulings landed from P2's diagnostics, both instances of the same theme — the converter was right and the model couldn't express what Apple actually draws:

• List-marker look: RIPE's bullets were invisible because our marker span was unstyled; P2's viewer fix (inherit the first run's style) makes them visible-white, but Apple colors markers from the list style's own font/color/scale. Landed markerColor/markerFontName/markerScale/markerBaselineOffsetPt on ListFormat — P2 extracts and renders, so those dots go Apple-orange.
• Shrink-to-fit: the ppd deck's "small title" mystery inverted — our 98pt is the stored truth; Apple shrinks it to ~81pt so it fits the placeholder in two lines. That's the same mechanism behind your truncated-text-boxes complaint, so I landed an explicit textFit: "grow" | "shrink" on text-hosting drawables (absent = fixed/clip) and handed extraction+rendering to the Keynote round-3 agent, which was already tasked with the truncation class — now it implements from a real flag instead of heuristics.

Also per P2: Home.key's title/subtitle proportions now measure matching Apple, and phantom borders on Keynote theme paragraphs are gone (a stroke definition alone isn't a border — the enable fields gate it). All three round agents are still working; the tree builds clean and the only red test is golden_g2, which sits deliberately in K3's re-sync path.

Strong idea — shapes/lines/vector art are the one TSD layer every format shares, and our sweeps keep finding drawable bugs by accident rather than by design. Two adjustments to make it work smoothly: the mining doesn't need a full agent (I can census the corpus JSON directly), and the specialist has to launch after Keynote round 3 lands, since it owns the same files (tsd.rs/drawables.rs/drawables.ts) and two agents in those files would collide. So: mining starts now, specialist launches on K3's completion with the curated list in hand.

The miner is running across all 964 fixtures now. It scores each document on the actual drawable surface: preset variety, explicit bézier paths, connection lines, groups, rotations, dashed strokes, line-end decorations (the arrowhead gap K2 flagged is still open), gradient fills, vector/TIFF media, and image masks — then ranks. That gives the specialist a curated set spanning all three formats instead of whatever a random sample happens to contain, and the per-format spread you predicted (Pages layout + Keynote heavy, some Numbers) will fall out of the ranking naturally.

The specialist's brief, once K3 lands and frees the drawable files: own the TSD layer end-to-end (tsd.rs/drawables.rs converter-side, drawables.ts viewer-side), work the top-ranked fixtures with visual_diff against the right app per extension, and chase class-level fixes — line-end glyph placement, preset coverage gaps, connection-line routing, dash/cap/join fidelity, gradient direction conventions, freehand groups — rather than per-doc patches. Known open items already queued for it: the arrowhead line-ends (data emitted, viewer doesn't place them) and the "hand-drawn" smart-stroke textures rendering as plain strokes.

I'll launch it with the ranked list as soon as the census and K3 both land — mining and round 3 are running in parallel, so no wall-clock lost.

The census delivered exactly the campaign fodder you predicted. Highlights of the curated list (saved for the specialist's brief):

• Keynote dominates, led by a genuine monster: one deck with 6,639 shapes (6,265 explicit bézier paths), 2,169 line-end decorations, 1,367 vector-media assets, 510 dashed strokes, and 902 groups — a stress test and breadth test in one file. Also curated: a 160-connection-line deck (routing/anchors), a 986-masked-images deck, and a charts+gradients mix.
• Pages contributes a 158-shape layout doc with 124 line-ends, plus a template family of three sibling docs heavy in groups and masks — nice for regression tripling.
• Numbers has thinner but real material: gradient-filled shapes on table sheets, grouped paths, and one doc that overlaps the tables agent's pre-BNC work (flagged so they don't collide — the specialist gets shapes only there).

The target file also carries the open class items from earlier rounds: unplaced arrowhead glyphs, smart-stroke textures, connection-line fidelity at scale, and group-nesting stress on the 902-group deck.

Keynote round 3 is still working (truncation/textFit, TIFF, geometry blowups); the moment its completion lands, the Drawables specialist launches with this list — same-file collision is the only thing holding it, and no wall-clock is being wasted since round 3 is itself drawables work.

Pages round 2 is complete — every task closed with Apple-verified evidence, all gates green. The notable finds beyond the planned work:

• Footnotes were never extracted from modern docs at all — anchors live in a different attribute table (table_footnote, keyed to U+000E) than the one round 1 read. G5's footnote now renders as an inline superscript with the body at the page bottom under Apple's separator rule.
• The three new model fields are extracted and rendered: "(a)"-style list surrounds match Apple in G5, two-column sections pack as real columns (proven on a crawl doc, since G5 turned out not to have any), and drop caps float correctly — G5 page 5's big "T" matches.
• RIPE's bullets are now Apple-orange at the list style's own scale — the full marker-look chain works, with run-style inheritance as fallback.
• Full-bleed cover pages (00C Textbook) render as positioned canvas layers instead of in-flow soup.
• Goldens were re-synced from a clean build and are green at HEAD.

I also promoted all three fixture build checklists into the repo (committed under fixtures/golden/): the G6 Keynote deck (13 slides now — I added textFit-overflow and connection-line slides from the round-3 findings), the G6 Numbers acid sheet, and a new G7 Pages mini-fixture scoped to exactly what G5 lacks — two-column section, footnote near a page boundary (needed for the one remaining footnote gap: page-space reservation during packing), a TOC, and a 2-line/2-char drop-cap variant to catch parameter handling. Build any of them when you feel like it and the harness takes over from there.

Still in flight: Keynote round 3 (truncation/textFit rendering, TIFF transcoding, geometry blowups) and Numbers round 2 (chart colors, v3 cells, Calendar heights). The Drawables specialist launches on K3's completion with the census-curated target list.

Fixed and committed — the harness no longer steals focus. The open -a calls activated the apps by default; open -g -a opens documents in the background, and the actual export/close chain is pure Apple events that never needed focus. Verified end-to-end on G1: open → export → rasterize → close, no activation. Both running agents pick this up automatically on their next visual_diff invocation since they invoke the script fresh each time.

So you may not need the second Mac after all: document windows will still appear and stack up in the background while agents verify, but your keyboard focus stays where you put it. (If even the window churn is too noisy, the other Mac remains the zen option — the checklists are committed in fixtures/golden/, so either machine works; just drop the finished .pages/.numbers/.key files into fixtures/golden/ and I'll wire them into the harness and re-sync expectations.)

And agreed on priorities — the reconciliation loop keeps running regardless: Keynote round 3 and Numbers round 2 are still grinding, with the Drawables specialist queued behind K3.

All set:

• Pushed the backlog: 49 accumulated commits went up to github.com/peterheb/pnk in one fast-forward, and both running agents now push after every commit — the public history tracks the work live from here on.
• Casing: nothing to fix — git already tracks AGENTS.md uppercase; your lowercase mention just resolved case-insensitively on APFS.
• CLAUDE.md created and pushed: it @AGENTS.md-imports the full project brief (so Claude Code sessions get it by default now), plus a tight Claude-scoped addendum capturing what this session proved out: the push-every-commit paper-trail policy, the build + visual_diff ground-truth loop (with the no-focus-steal behavior), the pre-commit gates, the golden guard, model stewardship with the "viewer never walks an inheritance chain" ruling, the per-agent file-ownership split, and the golden-fixture checklist workflow. Any future session — or parallel agent — starts with the same working rules without re-deriving them.

And glad the background activation is behaving! Keynote round 3 and Numbers round 2 are still grinding with their exports now invisible to you; the Drawables specialist launches the moment K3 reports.

Sorted, and no laundering needed — since all agents share one local repo, any successful push syncs the whole branch, so N2's commits (chart seriesColors and the v3 cell-storage decoder — that's census class 1, the 376-warning doc, closing) were already carried to origin by a later push from this session. I've told N2 to skip the push step and just keep committing; I push on every wake-up, which lately is every few minutes, so the public paper trail stays near-live.

One knob only you can turn, if you want it: agent subsessions get git push denied by their auto-mode permission classifier. It's fine as-is (I relay), but if you'd rather have agents push the moment they commit, you'd need to allowlist git push origin main in the project's permission settings — say the word and I'll set it up via the config skill.

Numbers round 2 delivered the deepest converter work of the day (already pushed — K3's pushes carried it):

• Cell-storage v3 decoded — the last big census class closed. The layout was reverse-engineered from scratch (flags-u64 gating a fixed field order); the worst doc went 376→2 warnings, the LED price list from 28→1 with product photos now rendering via cell image-fills.
• The v4 layout was subtly wrong everywhere, not just edge docs — it used fixed slots where the format actually uses two flag bitfields (leading presence bits plus trailing format-key counts). The rewrite was corpus-validated on 294k numeric cells across all 158 Numbers fixtures with zero misreads; the 911 Budget doc went from denormal garbage (1.24e-322) to digit-perfect currency values.
• Chart series colors extracted from the per-chart-type proto slots with theme/private style precedence — the Running Log pace line is Apple's exact #ff9e41 orange now.
• Calendar mystery solved: those ~60px week rows are stored nowhere — Apple auto-fits rows whose spacer cells contain literal newlines; our pre-line rendering reproduces it within a few px.

Bottom line for Numbers: only 10 corpus docs still carry any table-degraded warning, nearly all just the blanket pre-BNC notice. That family is essentially conquered.

K3 is the last one running (its recent pushes show TIFF→PNG transcoding landed and shrink-to-fit format docs in progress). When it reports, the Drawables specialist launches with the census-curated targets — the last planned campaign of the day.

K3 landed — all six classes fixed with real mechanisms, and 13 of the 15 sweep decks are now presentable. The truncation fix went three layers deep (real shrink_to_fit flags through the style chain, auto-grow boxes, and a bounded metric-drift pass for browsers' taller fonts); the LISA "geometry blowup" turned out to be shrink-to-fit all along; TIFFs transcode to PNG lazily in the wasm layer; and .ai backdrops now get their Flate streams inflated in-browser to find the page fill. Time for the finale — the cross-format Drawables specialist, now with exclusive run of the tree:

Status: the Drawables specialist is the only agent running, working the census-curated targets — line-end arrowheads, connection lines at scale, the monster deck (with perf numbers), K3's leftovers, group/rotation audits, Numbers sheet shapes. Everything else has landed and pushed.

The design ledger is now closed out and pushed too: every ruling from the model review is implemented — all three inheritance violations (resolved style chains, master underlay with strict paint-list membership, template underlays), both compactness leaks (aggregated warnings, truncated row/column arrays), and seven post-review model additions each with its extraction landed (grouping, numberSurround, section columns, footnote placement, drop caps, marker-look fields, textFit). Only minor type-hygiene nits remain open.

Where fidelity stands after today's five completed campaigns: Keynote — 13 of 15 sweep decks presentable, with truncation/shrink-to-fit, TIFF, ghost prompts, and backdrops all mechanically fixed; Numbers — v3/v4 cell storage fully decoded (294k cells validated, zero misreads), 10 docs left with any table warning; Pages — pagination with margins, page-layout underlays, footnotes, drop caps, columns, and the full G5 sweep. When the specialist reports I'll consolidate the final state, and the three build-checklists (G6 ×2, G7) are in fixtures/golden/ whenever a second Mac and a spare half-hour coincide.

Everything is landed, pushed (zero unpushed commits), and green. Here's the state of the world after the full campaign wave:

Drawables specialist (final campaign) — all seven items closed, zero model changes needed: - Arrowheads finally render — and the placement contract (glyph apex on the tip, scale = 0.4 + 0.6×stroke-width, head-at-end/tail-at-start, stroke cutback) was reverse-engineered from Apple's own PDF vectors and documented with provenance. - Connection lines rebake from live anchors — stored paths go stale when shapes move; the 160-connection deck now routes all its graph edges identically to Apple. Also killed a viewBox bug that was blowing hairline verticals into full-slide black bands. - The monster deck (181MB, 645 slides, 15,316 drawables): converts in 5.65s, first in-browser render 2.6s, slide switches under 100ms — no perf work needed. It still exposed two systemic fixes: textbox fills/borders now paint, and 0×0 textboxes anchor by paragraph alignment. - Per-axis path normalization (the white-band bug), reflections fading the right direction, negative marker indents un-clipped with Wingdings→Unicode stand-ins, page-layout canvases painting their resolved backgrounds (Labothek covers were white-on-white), and Numbers sheet shapes verified. The "legacy crop" leftover turned out not to reproduce, and smart-stroke stays a documented degrade rather than inventing fake wobble.

The day in total, six completed campaigns: Keynote went from "most corpus decks are disastrous" to 13/15 sweep decks presentable; Numbers has v3+v4 cell storage fully decoded (294k cells, zero misreads) and ~10 docs with any warning left; Pages paginates with margins, and G5's entire feature sweep — drop caps included — matches Apple. The model review's every ruling is implemented, the shell UI does whole-window drop with dark mode and JSON export, and the harness no longer steals your focus. All of it is on github.com/peterheb/pnk as small, explained commits.

Open threads, ranked: font-metric drift on tight line-spacing with fallback faces; in-browser PDF rasterization for vector media without raster twins; line-exact WP pagination; the three fixture checklists in fixtures/golden/ waiting for your other-Mac session. One stray: an agent left an untracked debug probe at crates/pnk2json/examples/ps_probe.rs — harmless, delete or keep as you like.

Both blank-line bugs fixed, verified against Apple's PDF, and pushed as 20b98bd.

They were actually two opposite bugs with the same symptom class:

• 1eb960ba (blank lines not blanking): the converter was emitting the empty paragraphs correctly (items: []), but the viewer rendered them as empty <p> elements, which collapse to zero height in HTML — so the gaps around the title and after the attendee list vanished. Fix: when a paragraph produces no visible content, append a <br> so the line box gets its strut at the inherited size. This also keeps blank lines measurable for WP pagination.

• G2 (extra blank lines): the reverse — no empty paragraphs exist in the archive at all. The "Shapes can hold text too." box and the cat-photo caption are tight two-paragraph runs with no pooled paragraph style, so applyParaStyle never ran and the browser's default 1em <p> margin showed up as a phantom blank line. The WP print area already had a margin: 0 reset with the rationale "Apple has no implicit paragraph margins" — shape/textbox text (.drawable-text) just never got the same rule. It does now.

Both composites re-verified by eye: 1eb960ba's gaps are back where Apple puts them, and G2's two text boxes now match Apple line-for-line. Viewer gate 6/6, no converter changes so the golden expectations are untouched.

Yes — :8123 is the esbuild static server over viewer/dist/, which I rebuilt at 11:06 before the last verification pass, and I just confirmed the live responses contain both fixes (the "br, img" strut check in main.js and the new margin reset in styles.css). Just hard-refresh the tab (⌘⇧R) in case your browser cached the old bundle.

Fixed and pushed as 5daf1d3. You were right that it was still excessive — the margin fix was only half the story.

The remaining leading came from the line-box strut: the run spans carry their own font sizes (9pt for that caption), but the <p> block itself was inheriting the app chrome's 15px. Every line box is at least as tall as the block's own font metrics demand, so small text got ~18px line pitch — which is why the gap appeared even between the two wrapped lines inside the URL paragraph, where margins couldn't be the cause.

The fix sizes the paragraph block to its largest visible run, which matches Apple's own rule (line height derives from the tallest run in the line). Verified by eye against Apple's PDF: the caption now sits at Apple's tight three-line leading, the heading box matches, 1eb960ba is unchanged, and the dns Keynote deck renders pixel-identical to its pre-change sweep composite (big text already exceeded the old strut, so it never noticed). Gate 6/6. Hard-refresh your :8123 tab to pick up the new bundle.

One unrelated thing I noticed in that Keynote composite: Apple's PDF export omits slides marked "skipped" while our viewer shows all slides, so side-by-side pages drift out of register on decks with skipped slides. Harmless for the viewer itself (showing skipped slides is arguably a feature), but say the word if you'd rather have the viewer badge or hide them.

Fixed and pushed as 01021ad. Good catch — and your instinct about what was happening was exactly right, just with the causality inverted: the green wasn't the diamond's shadow, it was the pentagon's reflection being painted a full shape-height too low, so only its far end peeked out from under the diamond while the contact line got nothing.

The story: the converter was blameless (pentagon correctly owns both the shadow and reflection in the JSON). The viewer bug was in the -webkit-box-reflect mask — Chromium flips the gradient mask together with the mirrored copy, so the gradient's top edge lands on the far edge of the reflection, not the contact line. The Drawables specialist assumed the opposite yesterday (commit 43879e8) and inverted what had been a correct ramp. Rather than guess again, I wrote a three-box probe page and screenshotted it in Chromium 151 to pin down the semantics, then restored the transparent→opaque direction while keeping the specialist's measured Apple geometry (full strength at contact, dead ~55% out).

Verified by eye on both sides of the dispute: G2's pentagon now shows the green wash right at its base with the diamond painting over it (drop shadow intact), and the 3c844ac1 logos deck — the very fixture the specialist cited — now matches Apple's export too, with IIT/Tarides mirroring at contact and fading down. Gate 6/6. Hard-refresh :8123 to see it.

Solid list — several of these look like shared root causes (the notes leading smells like the same UA-margin leak I fixed for shapes, and the loose line spacing is probably line-height: normal vs Apple's tighter default). I'll run ground truth on the deck first, then work through the fidelity items, then rebuild the Keynote UI as a continuous scroll.

All seven fidelity items plus the UI rework are done, verified against Apple's PDF, and pushed as three commits (676533d, f8bbc68, bc5004d). The interesting finds:

Slide 1 title (and 27/28's quotes) — a format archaeology dig. This is a Keynote-'09-converted deck (format 1.5), and its text shapes store geometry with flags=0 where position means the shape's center — the title stores x=512, the slide's exact midpoint. Modern archives always write flags=3 (verified across G2 and other decks, plus this deck's own correctly-placed images). The converter now re-anchors these to top-left, so the viewer never learns the flag exists. The flag bits are documented in no local proto, so the commit carries the full evidence chain.

Slide 9 — a real bug hiding under "not centered." The shrink-to-fit pass compensates its scale with a >100% width, but the inner div is a flex item and default flex-shrink:1 silently squashed it back — text wrapped at the wrong width, so centered lines drifted left ~45pt. One flex: 0 0 auto fixes every shrink box, not just this slide.

Slide 2/19 arrows: the bullets are markerKind: "image" (a rightArrow PNG in the media store) — the viewer only knew string/number markers. They now render scaled like glyphs, with their indents.

Slide 19's loose leading: the raised red labels (baseline-shifted runs) were growing each 40px-exact line box to 41.7px. Shifted spans now get line-height: 0, Apple's own rule (baseline shifts never affect leading).

Slide 29's quote: a 0×0 shape — it now gets the same point-anchor natural-width layout the 0×0 textbox labels already had, so it's one 51pt line instead of a four-line sliver.

Thumbnails: the frame's 1px border eats into the border-box, so every canvas was scaled 2px too wide and clipped right/bottom by overflow:hidden — ~12 doc-px at thumbnail scale.

UI: slides now stack in one continuous scroll like Pages, lazily rendered (aspect-ratio placeholders + IntersectionObserver, so the 645-slide monster deck stays fast). The thumbnail rail is sticky and clicking scrolls to the slide. Notes are a compact 13px strip under a slide only when it has visible notes — empty note storages (the gate deck had 11 of them) render nothing. That also killed the slide-10 leading problem. The visual_diff harness now screenshots per-slide canvas frames, so composites align canvas-to-canvas.

Gates: conformance NO DEFECTS, goldens byte-stable, cargo tests green, viewer 6/6.

Done and pushed as 95c7ada. The viewer is now pnk.vu:

• Top nav bar (slim, sticky, blurred): monospace pnk.vu wordmark with the accent-red .vu, then the filename, a quiet app badge — colored text chips (pages orange / numbers green / keynote blue) replacing the old red pill — compact format/locale meta, and right-aligned json / open… actions. The brand is the way home; open… goes straight to the picker.
• Fonts list removed entirely — it was inspector chrome, not viewer UI.
• Warnings demoted to a whisper: a small amber ▲ n in the nav that exists only when the decode actually produced warnings; clicking it drops down the detail rows. Clean decodes show nothing.
• Landing card re-branded to match: mono lowercase wordmark, terminal-flavored tagline, dot-labeled .pages / .numbers / .key chips, tighter spacing.
• Everything sits a bit more compactly (slimmer paddings, wider 1160px content column), and the Keynote rail's sticky offset tracks the new nav height.

Verified by screenshot in light and dark mode across landing, document, and warnings-dropdown states; tsc clean and Playwright gate 6/6.

One taste note you might want to weigh in on: the wordmark renders as pnk.vu in SF Mono with the dot getting a full monospace cell — I kept it because the airiness reads intentionally terminal-ish, but it's a one-line change if you'd rather tighten the dot.

Pushed as a85b362. The landing is now one sentence — "view iWork files in your browser. nothing is uploaded." — with the chips, the drop zone, and a six-word hint. The nav open… button only exists while a document or error is showing, so the accent-red "choose a file…" is the lone CTA dead-center on the landing. Gate 6/6 with the hint assertion updated.

Done, pushed as f72882d. The nav json button now toggles a pretty-printed, syntax-colored view of the model — purple keys, green strings, blue numbers, with dark-mode variants — and the button shows an active (inverted) state while it's open. The toolbar above the code shows the filename and pretty-printed size, and holds the download button (same local blob URL, no network).

Two engineering details worth knowing: the rendered document DOM stays alive underneath, so toggling back is instant even on a big deck; and syntax coloring caps at 3 MB of pretty text — a dense Numbers envelope can hit tens of MB, where a span-per-token DOM would hang the tab, so those fall back to plain (still pretty-printed) text. Render is lazy on first toggle and cached per document.

All three shipped as 38c264c:

• Footer: the Apple trademark line ("Pages, Numbers, Keynote and iWork are trademarks of Apple Inc. pnk.vu is not affiliated with or endorsed by Apple.") plus a collapsed privacy & terms toggle with the combined blurb — runs entirely in-browser, no personal data, Cloudflare keeps standard server-side logs and anonymous analytics, provided as-is, MIT/Apache-2.0 with a GitHub link.
• Nav: a quiet monospace github link to the repo, muted until hover.
• Smarter pretty-printing: the JSON view now keeps any object or array on one line when its inline form fits a 76-character budget — so "position": { "x": 511.5, "y": 728.5 } and "pageSize": { "width": 612, "height": 792 } read like a person wrote them. It's built bottom-up (each node stringified exactly once), so multi-MB envelopes pay no quadratic cost, and dense grid rows collapsing to single lines actually shrinks the pretty output substantially.

Gate 6/6 — the github anchor doesn't trip the zero-network assertion since it only fires on click.

One thing I noticed while verifying the JSON view on G2: the envelope has a hiddenBody carrying the text "[where is this mystery text]" — that's text you typed somewhere in the fixture that isn't rendered on any page (Pages word-processing bodies exist even in page-layout docs). Might be worth a look at whether the viewer should surface hidden bodies somehow, or whether it's fine as a model-only artifact.

All four done, pushed as a95c79d + 09cb423:

• Hint removed — the landing is down to the wordmark, one sentence, chips, and the drop zone. (#drop-hint survives invisibly as the parser-load-failure slot.)
• Octocat — the nav github link is now the mark itself, muted until hover.
• Warnings dropdown — proper padding, and a divider between the per-code summary chips and the detail rows.
• Numbers borders — real converter bug, confirmed and fixed. Your "sometimes" was exactly right: the default gridlines are emitted half-open (each cell owns its right+bottom edge, so sidecar border-erases keep a single owner per edge) — which left the table's top and left outer boundary owned by nobody. A corpus scan found dozens of docs with zero top borders on row 0 while every interior boundary was stroked. Apple draws the frame with dedicated strokes I hadn't been reading — table_body_*_border_stroke (f58/f59, gated by table_border_visible) plus header-region border strokes (f47/f50). Those now bake onto boundary cells as the weakest overlay, so explicit sidecar strokes and erases still win. Bottom/right outer edges also now use the true frame stroke instead of borrowing the interior gridline.

Verified: corpus rescan shows zero tables left with unowned outer edges, spot-checked by eye against Apple's PDF on two flagged docs, goldens byte-stable, conformance NO DEFECTS, gate 6/6.

One thing I noticed en route, for the backlog: Apple's PDF export print-paginates wide Numbers sheets into page columns, so visual_diff composites for such docs pair one long viewer screenshot against many narrow pages — misleading but harmless. A future harness tweak could tile our screenshot to Apple's print geometry.

Pushed as 6985d83. On screens under 700px:

• Keynote: the thumbnail rail is gone — the continuous scroll is the navigation, and slides run full-width.
• Numbers: sheet tabs stay on a single row and scroll horizontally instead of wrapping into a stack; sheet-area padding tightens.
• Pages: page frames no longer hardcode a 720px display width — they now size to min(720, container), so documents fill a phone edge-to-edge. This is display-only scaling: pagination still measures in page points, so page breaks are identical at every screen size. Desktop is unchanged (container > 720).

I verified all three by screenshot at 390×844 — the Spamhaus deck scrolls cleanly slide-by-slide, the budget spreadsheet shows its tab strip and bordered table properly, and the ATM word-processing doc reads full-width. Also caught and removed a stale CSS rule from the old warnings-panel era that was silently overriding the new dropdown's padding.

One caveat worth knowing: the mobile width is measured at render time, so rotating a phone (or resizing a window across the breakpoint) after a document is open keeps the old page scale until the file is reopened. Fine for a hackathon demo; a resize listener is a small follow-up if you want it.

Done, pushed as bf44ba7. The README is now a tight one-screen read: what it is, feature list matching what actually shipped, verified quick-start commands (checked against the real npm scripts and build script), repo layout table, a four-line format primer, the four verification mechanisms, and license + Apple trademark note. It links to pnk.live as the live site. docs/model-review.md got a one-line freeze stamp marking the §2 hygiene nits as post-hackathon carry-over; everything else in docs/ was already factual and dated, so I left it alone.

Two things to know for the deploy from the other machine:

• The final build artifact is viewer/dist/ — regenerate it there with cd viewer && npm install && npm run build (needs rust + wasm32-unknown-unknown + wasm-bindgen 0.2.127), or just rsync the dist/ from this machine since it's fully static.
• The in-app branding still says pnk.vu (wordmark, tab title) while the site will live at pnk.live. If that's intentional, ignore this; if you'd rather the wordmark match the domain, it's a two-file find/replace — say the word before you build for deploy.

Fixed and pushed as dd5d81b — README now points to pnk.vu, which also means the in-app branding matches the domain perfectly. Nothing to change before you build for deploy.

The tree is fully synced: every commit is on origin/main, all gates green as of the freeze. The only untracked leftovers are the Apple template binaries (deliberately untracked) and that stray debug probe in crates/pnk2json/examples/ — neither ships.

It was a genuinely fun build — from Snappy blocks to arrowhead geometry to a product with a wordmark, in 48 hours, with a public paper trail of 200+ commits to show for it. Good luck with the video and the voting — go get 'em! 🎉