Record the sweep that leaves no word specimen incomplete

[MarkusNeusinger]

The words.json sidecar defines incomplete for a word specimen whose OWN
ink the crop rect clips — the flag that takes a never-traceable specimen out
of the admin's „Offen" list and out of the tracing tally's denominator (#470).
Not one of the 202 entries carries it. This PR checks whether that is right,
and writes the answer down where the flag is defined.

What was measured

A boundary sweep over every rect, three sensors, all read-only:

• Ink on the border line of each rect, excludes painted white first.
• Connected components on the whole plate, so a stroke that leaves a box
  is one object seen from both sides. This is the sensor that matters: a
  per-crop adaptive threshold fringes on the abb22 scan and reports dozens of
  phantom edge touches.
• Diacritic-sized specks in the band above a rect, for the failure the
  edge scan cannot see: an i-dot left outside with a gap, touching nothing.

Result: no specimen is missing a letterform or a diacritic. No body ink
leaves any rect to the right, the top or the bottom at any threshold down to
0.78 grey; the only own ink outside a rect anywhere is a lead-in hairline,
at most 19 px. No orphan diacritic sits above any rect. Every rect keeps the
3 px BOX_PAD_PX standard except where foreign ink reaches in.

The eleven rects an ink component does cross:

Specimen	What crosses	Verdict
Gewehr, Zügel, streiten, a22-dank	the comma after the word	foreign — same case regieren, a22-Krüger, a22-dank-2, a22-hinaus already carry as an exclude
laden, a22-laden	the descender of the line above	foreign
a22-fern, a22-ein	the previous word's last letter / run-out	foreign
a22-laden, a22-ihren, a22-doch, a22-roten-2	the word's own lead-in hairline, ≤ 19 px	intact — the letterforms are whole and the renderer generates the Anstrich anyway

A blind second pass over the rendered crops, given no expected answer, agreed
on all sixteen specimens it was shown.

Why the answer is a negative, and expected

The seven specimens that really were clipped — regieren, zum, einer,
das, und, Wer, zwei — were repaired on sep01 and the word bench
re-baselined (qualitaetsmetrik.md §15), because the flag is the second
remedy, not the first: where only the RECT cuts the ink, the rect is fixed
(tools/wordbench/repair_boxes.py), and incomplete is left for ink that
ends on the plate itself. Six of those seven are in the candidate list an
automatic edge scan still produces — the list predates the repair.

So the sidecar note now carries both halves: the repair-before-flag rule, and
the sweep with the entries it clears by name. The next edge scan that lists
the same rects then reads as a settled question rather than as work waiting.

What is NOT in this PR, and why

• No rect changed. The corners are frozen wordbench fixtures; widening one
  is a declared re-baseline, and nothing here needs one.
• No exclude added for the four commas that reach into their crop,
  although regieren's comma has exactly such a box. Adding one changes
  ref_mask and is therefore a re-baseline too — the author's call, noted
  here as a finding rather than taken.
• No changelog fragment: data-only PR, exempt per changelog.d/README.md;
  the provenance lives in SOURCE.md, whose Maße/SHA256 for words.json
  are updated in the same commit.

Verification

• Diff is the sidecar's top-level note line plus the SOURCE.md block — no
  entry touched, so no rect, exclude or lineature moved and the frozen
  fixtures stay byte-identical.
• uv run --extra test pytest — 2465 passed, 18 skipped.
• ruff check + ruff format --check — clean.
• npm run test — 296 passed (the incomplete filter path is pinned by
  traceStatus.test.ts and is untouched).
• License battery for the scoped diff: all 15 recorded SHA256 hashes match
  their bytes, words.json included; required SOURCE.md fields intact; no
  new files.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UEScQMZFvxxNNyNJYryfa3

[Copilot]

🟡 Changes recommended

Clarify the incomplete rule for clipped lead-ins and correct the provenance claim about seven clipped crops.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Documents the audit confirming all 202 word specimens remain traceable.

Changes:

• Records boundary-sweep results and the repair-before-flag policy.
• Updates integrity metadata and audit provenance.

File summaries

File	Description
data/sources/suetterlin-1922/words.json	Adds completeness-audit findings and policy notes.
data/sources/suetterlin-1922/SOURCE.md	Updates checksum, size, and audit provenance.

Review details

Suppressed comments (1)

data/sources/suetterlin-1922/SOURCE.md:74

• The cited §15 says seven crops were below the 3 px standard, but only four had negative clearance; the other three still had 1–2 px of air (qualitaetsmetrik.md:3646-3659). Calling all seven “wirklich angeschnitten” makes this provenance note contradict its source.

             Buchstabe oder ein Diakritikum fehlt. Die sieben wirklich
             angeschnittenen Proben sind seit der Rechteck-Reparatur
             `sep01` repariert statt markiert
             (`qualitaetsmetrik.md` §15) — die Marke ist das zweite

• Files reviewed: 2/2 changed files
• Comments generated: 1
• Review effort level: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.