[DemchaAV]

Why

The editing protocol asks of every floating drawing whether it followed the text it stands beside
when an edit moved that text. A drawing nailed to the sheet cannot: wp:positionV relativeFrom="page" places it from the top of the page it is drawn on and from nothing else, so the
text beside it has no claim on where it stands. Measured on the self-check's two alike probe
documents, with the paragraph above lengthened by two lines: such a drawing moved 0.00pt in Word
16 and 0.00pt in LibreOffice 24.2 while its own text moved 32.04 and 52.50pt, where the same
drawing carried by that text's paragraph in the other document moved 31.96 and 52.50pt with it.
Both editors, by construction.

Counting that as a drawing which failed to follow measures which anchor the export chose, not
whether the export holds a drawing to its text — and which anchor the export chose has an instrument
of its own. DocxDrawingAnchorGuardTest names every drawing of the corpus still placed from the
page, one by one with the reason it is there, and fails when one is added to that list or taken off
it. So the figure this protocol reports was made of two questions at once, and the one it is asked
for could not be read off it.

What changed

scripts/docx-visual/edit-protocol-corpus.ps1

• What a drawing is placed down from is read from Word, per drawing. Get-DrawingPlaces carries
  RelativeVerticalPosition beside the place and the left edge. Measured over the corpus both ways
  round: the files hold 774 wp:anchor in the body and Word names 774 shapes there, with no anchor
  the one reading has and the other has not, and wp:positionV relativeFrom agrees with Word on
  every one — 46 from the page, 14 of them page fills, and 728 from their paragraph, with no third
  answer and no name carried twice. The 32 drawings left are the 32 the guard's ledger names,
  document for document and name for name (invoice-payments 13, invoice-workspace 8,
  invoice-merchant 4, invoice-luma_studio_long 6, invoice-obsidian_long 1). The 18 further
  anchors the files hold outside the body are page fills in headers, which this reading, being of
  the main story, does not reach.
• The nailed drawings are measured as the carried ones are, named with what they did, and left out
  of the verdict. Measure-Following is the one measuring of a set — how many stood beside text
  the edit moved and moved with it, how many stood beside text it left standing and stood still —
  and it is run over both sets. The detail names every nailed drawing of the document, with the
  counts for the set and by name for each one that left its text or lost its place. Nothing is set
  aside quietly: the number is printed, the names are printed, and what the set did is printed.
• A drawing that ought to follow stays in the verdict whether or not nailed drawings stand beside
  it, since what is read is each drawing's own anchor and not the document it stands in. The
  corpus shows it: invoice-merchant sets its four nailed drawings aside and still fails, on
  Drawing 15, which a paragraph carries and which crosses onto the page after while the text it
  stands beside stays.
• Nothing may leave the verdict by carrying an answer the scenario has no question for. A
  drawing placed down from anything that is neither its paragraph nor the page fails with its own
  reason — the export writes only those two (DocxDrawings and DocxPageBackgrounds are the only
  writers of wp:positionV relativeFrom) and the corpus holds nothing else, so this is a hole
  closed rather than a case seen. The same hole was already open for a name Word gives to more than
  one drawing, which Get-DrawingPlaces reads as no drawing at all and the scenario passed over in
  silence: such a name is now named and fails, and is reported on a line of its own rather than
  among the places the page geometry does not give, since the geometry is not what failed there and
  a reason naming the wrong cause sends the next reader after a defect that is not there. It
  predates this change, and is in scope because this change makes the principle explicit in three
  files.
• A drawing the reading no longer finds is counted apart from one whose own text cannot be placed
  again, and fails for either set, of the drawings each set measures. The two were one count
  before. The first — Word does not name the drawing after the save and the reopen, or names it and
  another alike so neither can be told from the other — is a drawing gone from the document or one
  that can no longer be told apart, which is not a question of what it was placed from, so taking
  the nailed drawings out of the verdict must not take that with them. The second is the instrument
  losing the paragraph a drawing was held to, which answers nothing about a drawing nothing is asked
  of, so it fails for the carried set alone. Both are nought over the corpus.
• A document whose drawings are all nailed is measured rather than recorded as having nothing to
  edit, which would say something untrue of it; its detail names them beside the nought carried
  drawings held to the verdict.

CONTRIBUTING.md says what the scenario asks of whom and what holds the rest; CHANGELOG.md
carries it under Tests. The export is untouched, so docs/recipes/docx-export.md and the capability
matrix are unchanged, and no public API moved, so the knowledge pack is unchanged.

Verification

• Over the template corpus, 71 documents, run five times across the builds of this change,
  agreeing exactly on the status and all five counts of every one of the 71 every time:
  • 305 of the 320 drawings a paragraph carries standing beside text the edit moved follow it —
    95.3% — and 400 of the 408 beside text it leaves standing stand still too.
  • The 32 nailed to the sheet, measured the same way: 13 stood beside text that moved and none
    followed it, 19 stood beside text that stayed and all 19 stood still.
  • Added back together those are 305 of 333, 91.6%, and 419 of 427 — the two figures this
    protocol last reported, to the drawing. So the denominator is all that changed, and the run is
    comparable to that one count for count rather than by assertion.
  • 30 documents pass, 8 fail, 33 have nothing to measure, where it was 28, 10 and 33. The two
    that pass now failed only on nailed drawings: invoice-payments, whose eleven line-item tiles
    do not move while their text moves 40pt, and invoice-luma_studio_long, two of its six. No
    other document's verdict changed.
• -SelfCheck: 15 of 15, where it ran 10. One probe changed and five are new: the drawing
  nailed to the sheet at the place it already stands, now held to passing and to naming what it did;
  a document holding a nailed drawing beside a carried one that leaves its text, where the carried
  one must still fail; a drawing nailed to a page with no text on it, which is the one disposition
  the corpus does not reach; a nailed drawing whose place the geometry cannot give, paired with the
  document where a carried drawing fails for that same reason; a drawing placed down from the
  page's margin; and two drawings sharing a name. The last three are written here because nothing
  the export produces reaches those cases.
  • A paragraph holding only a page break was measured and refused as the "no text on its page"
    probe: the break is a character of that paragraph's text, so Word places the paragraph on the
    page before the break and draws a drawing nailed from it on that page too — where the probe
    found the body's own paragraph to stand beside. A tall untexted row pushed onto the page after
    reaches the case.
• Thirteen sabotages, nine caught by the probe written for each, the file restored byte for
  byte from a copy taken at the start of each run and asserted against it by sha256:
  • a nailed drawing let into the verdict again → the nailed probe fails;
  • what it is placed down from no longer read from Word → the same probe fails;
  • the nailed drawings no longer named in the detail → the same probe fails on its reason;
  • the ones that left their text no longer named one by one → the same probe fails on its reason;
  • the nailed drawings counted but no longer measured → the same probe fails on its reason;
  • every drawing of a document holding a nailed one let off → the mixed document passes, which
    is the false pass this change had to be written not to buy;
  • anything but a paragraph taken for nailed → the margin probe's document passes where it must
    fail, so that probe fails;
  • an unreadable nailed drawing failing the scenario again → the paired probe fails;
  • a shared name passed over in silence again → the two-of-one-name probe's document passes where
    it must fail, so that probe fails.
  • Four more are reached by nothing, and are run to measure that rather than to pass. A
    drawing the reading no longer finds no longer failing, once on the carried side and once on the
    nailed one: no document can make Word drop a shape or give two shapes one name over an edit. And
    the two exclusions the nailed set carries of its own — a drawing whose own text cannot be placed
    again, and one whose place the geometry stops giving after the edit — which are counted and
    reported and fail nothing, rightly by this change's own reasoning, but which no probe reaches
    either. All four were broken on purpose and nothing caught them.
  • Two more things go unprobed and are named rather than claimed away: the order the names print
    in, which no probe has two entries to distinguish and which changes no verdict; and one path out
    of the verdict older than this change — a carried drawing whose page holds no paragraph the
    reading could place is counted as standing beside no text, and a document can pass with it. The
    corpus reaches neither.
    Each run is read for the probe's own name among the failures and for the count line saying 15
    checks, so a sabotage that stopped the script before it measured anything cannot pass for one that
    was caught.
• The export is untouched: the corpus was exported again from this tree and compared with the
  export the measurements were taken on — 71 documents, no part changed, the report's notes 1497
  either way.
• ./mvnw clean verify over the reactor, the four knowledge checks, core javadoc at 0 warnings,
  and the examples suite (93 tests, CommittedAssetDriftTest among them): all pass on the tree as
  it ships. The documentation guards that read CONTRIBUTING.md were run on their own as well — 24
  tests, 0 failures.

What the figure is and is not: it is taken over the drawings the export already anchors in a
paragraph, so it says how well those keep to their text and nothing about the 32 the ledger still
holds — and that ledger is of work to do, not of cases where the sheet is right. The eight failing
documents answer for 23 carried drawings: 15 that do not follow text that moved and 8 that move from
text that did not. The margin is thin and worth knowing before the figure is quoted as settled: each
further carried stray costs about a third of a percentage point, and 304 of 320 is exactly 95 in a
hundred.

The protocol needs Word and is run by hand, so no CI job covers it; -SelfCheck is what to run
after changing how a place is read, how an owner is chosen, or what the verdict is taken over.