# Shared table typography and complete static exports

The shared table repair passes current automated verification, including the
final 421-slide bank's observed exit zero and all twelve checks. `state.json`
binds the delivered artifacts and retained evidence. This is not completion of
the exhaustive presentation project or the eight craft standards in `TODOS.md`.

Baseline: `6c3186d2a4bac18f17acdbb63b3177759f32d6df`. Source pin:
`1ceae52c7896a59ac74a4429d410f3218f0b91c6`. The library remains 40 guides, 421
unique slides, 447 guide instances and 287 cited sources.

## Changes and preservation

The [implementation brief](../../authoring/table-typography.md) applies the
mandatory frontend skill's typography, restraint and actual-size review to the
existing cream/ink/terracotta comparison plane. No decorative imagery or new
motion is introduced.

- Ordinary slide-table headers now share the existing 16px quantitative-header
  floor, authored case and normal letter spacing, including compact playback.
  The 40 delivered PDFs have 114 table instances and 322 headers. Poppler
  reports a 12px effective header size at 960px page fit for every header, using
  embedded Unicode-mapped Liberation Sans. Integer point metadata is not precise
  glyph height or a reader-approved threshold.
- Recognized number/unit expressions stay together without changing characters,
  signs, punctuation, units or values. Narrow row widths retain readable value
  columns and sticky row identity. Exact paths and optional identifier breaks
  remain separate from quantity recognition.
- A print scrolling wrapper could shrink and clip final-row ink even while all
  text remained extractable. Print now retains natural table height and visible
  overflow. The full browser/export guards check actual 1280px print geometry,
  every text fragment, page containment and separation from surrounding content.
- Source excerpts retain their independent 15px text and horizontal scrolling,
  without slide-table sticky positioning. A complete revision hash wraps inside
  the narrow dialog. Source cells keep a 160px minimum width so ordinary headers
  and row labels do not collapse into short fragments.

All authored slides, drafts, assignments, cited source-template bodies, search
and audio metadata, speech scripts and 421 recording hashes are unchanged. Exact
columns and rows are newly included in table export coverage. This changes one
dependent `citationMapSha256` field in the repository-discovery index; all 44
other discovery files and every inventory record remain unchanged. The
preservation report distinguishes 336 unchanged inputs from that one derived
binding. No source units are newly assigned or native/model pipeline executed.

## Automated evidence

Run all expensive targets serially with the repository resource supervisor. Logs
retain RAM, swap and disk observations; unrelated PID-prefixed host-process rows
are removed from the published copies; trailing whitespace and blank EOF lines
are normalized. Raw and retained hashes, omission counts and formatting changes
are recorded. An exit record, not an early `completed` field, is the terminal
evidence. Both complete current browser suites have observed exit zero.

| Check               | Scope                                                                                                                                                                                                                                                          |
| ------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Full bank           | All 421 slides at desktop/compact/mobile, accessibility, sources/transcript/navigation, 421 PDF pages, 109 print tables, real per-track playback/end transitions, HTTP failure/race handling and final artifact stability; all twelve checks, no subset flags. |
| Full center         | All 40 guides, 894 embedded captures, 80 real audio boundaries, 40 fresh PDFs and 114 print table instances; no subset flags.                                                                                                                                  |
| Table typography    | 1,180 checks and 310 captures: all 109 native tables across ordinary, compact-playing, mobile, narrow, enlarged and font contexts, plus actual 200%/400% tab zoom and focused embedded contexts; ten rejected defects.                                         |
| Complete print      | All 109 unique tables plus the delivered Metis specimen; five rejected defects and complete/clipped PDF contrast.                                                                                                                                              |
| Source dialog       | Real Tara disclosure, scrolling, exact cells/revision, axe and focus return at three widths; 15 inspected captures and five rejected defects. All 482 embedded source-table specimens are also isolated at each width, visiting all 4,431 column positions.    |
| Delivered PDF       | All 447 pages rasterized, all 114 table pages at both 960px and 1440px; exact table text/order, effective header size and actual font metadata across all 40 PDFs. An actual 11px-header PDF is rejected.                                                      |
| Focused regressions | Companion labels; contracts; table and composition pilots; the 168-check integrated connection-quality matrix and 60-check frozen draft matrix.                                                                                                                |
| Tooling and audio   | 226 Python tests, 274 Node tests, Ruff, ESLint; all 421 audio files checked for fingerprints, hashes, durations, decoding and encoded pauses.                                                                                                                  |
| Freshness           | Builder/preview, documentation coverage, repository discovery, workflow inventory and workflow coverage. Complete-teaching gates deliberately return one for the unchanged backlogs, not a passing complete claim.                                             |

No product API, persistence, mobile application or native code changes are made.
Historical adapted-host native results are not relabeled as new executions.
Full-workspace Nx, app builds and packaging are outside this renderer change.

## Review boundaries

The final screen captures exactly match the 310 previously inspected captures,
so observations transfer only by PNG equality. Initial, vertically scrolled and
last-column views remain separately labeled. At 400% zoom, the center's header
occupies substantial height and its selected title is truncated; this remains an
open navigation/accommodation issue. Isolated source filename fragments and
whole-slide initial-viewport fit also remain outside this repair's acceptance.
The explicit missing-font probe resolves to Liberation Sans on this host, as
does the preferred stack: this is not evidence for a distinct fallback face.

The final PDF comparison distinguishes three forms of evidence:

- Eighteen captures of nine repaired pages exactly match directly inspected
  repaired-PDF probes. These supersede earlier observations that missed
  clipping.
- Twelve current captures of six additional pages are directly inspected at both
  sizes, including the largest remaining pixel change and mixed prose,
  exact-source and quantitative comparisons.
- The remaining 198 captures inherit only their exactly unchanged raster
  regions. For all 210 non-repair image pairs, 3–13 pixels differ, exclusively
  at outer table edges: restored rule-end ink and, at 1440px, antialiased
  first-glyph edges. Every changed pixel is retained. The checker reconstructs
  the prior image in memory and requires equality at every other pixel; it
  rejects interior mutations, lost edge ink and missing difference records.
  These are not whole-PNG equality or new direct-inspection claims.

All 333 non-table pages have exact baseline raster equality. That establishes
unchanged output, not a new every-page optical review. The 15 current source
dialog captures are directly inspected; their isolated 482-table inventory
checks establish geometry, not complete source-excerpt flow or optical review.
Three current focused print images exactly match their inspected positive and
negative specimens. Historical weak alternatives remain clearly rejected.

`final-spot-review.json` records three supplemental direct inspections of the
current mobile score table and 960px Metis dimensions/network measurements
pages. Their hashes bind to the same retained images; they corroborate the
bounded findings without adding specimens or changing the original review
categories.

The retained images, reports, review ledgers, unsuccessful attempts, original
repair-input PDFs and prior edge-comparison rasters carry individual hashes. The
receipt verifier checks integrity and scope; the pixel verifier rechecks the
actual retained before/after image bytes without changing them:

```sh
node docs/presentations/presentation-center/verification/shared-table-typography-2026-09-07/scripts/verify-retained.mjs
python3 docs/presentations/presentation-center/verification/shared-table-typography-2026-09-07/scripts/verify-final-pdf-edge-inheritance.py --receipt docs/presentations/presentation-center/verification/shared-table-typography-2026-09-07
```

The pixel verifier requires Pillow. Browser reproduction uses the existing
Playwright/Chromium and axe dependencies; PDF checks use Poppler. Preserve fresh
output directories and use one supervised target at a time. Archived run scripts
retain the original temporary paths; the two receipt verifiers above are
portable within this checkout. Integrity is not a replacement for fresh browser
execution.

For fresh browser evidence, first run the builder with `--preview` to create the
ignored whole-bank fixture. Use the prepared presentation Python environment,
Playwright Chromium and Poppler described in `TODOS.md`. Review host resources
before each command and run these targets separately; each rejects reused output
directories and does not accept subset flags:

```sh
table_qa_root="$(mktemp -d "${TMPDIR:-/tmp}/oshun-table-qa.XXXXXX")"

OSHUN_TYPOGRAPHY_QA_DIR="$table_qa_root/typography" \
  python tools/presentations/run-supervised.py node tools/presentations/tests/table-typography.mjs

OSHUN_SOURCE_TABLE_QA_DIR="$table_qa_root/source-dialog" \
  python tools/presentations/run-supervised.py node tools/presentations/tests/table-source-dialog.mjs

OSHUN_PRINT_TABLE_QA_DIR="$table_qa_root/print" \
  python tools/presentations/run-supervised.py node tools/presentations/tests/table-print-regression.mjs
```

The typography target includes actual tab zoom through a disposable Chromium
profile and the local test-only extension. The full standalone and center
commands in `TODOS.md` also exercise the shared table and print guards. Inspect
fresh screenshots and PDF rasters at their recorded sizes before attaching new
review findings; regenerating output does not inherit inspection automatically.

Normal-speed listening, willing-reader tasks/accommodations/retests, accepted
benchmarks, complete typography/optical specification and every-slide craft
review remain open. So do exhaustive source teaching, actual Eve integration, ad
hoc narrated artifacts, and the final Domains, diagram and UI-design passes.
