Various diff fixes needed

[pmolodo]

For now, just creating a list of issues I found, will separate out into individual issues later.

When inspecting aousd_core_spec.diff_b7bc21a_to_e98ceb3.pdf, a number of issues were found:

(Note - section numbers and page numbers reference the above pdf - they may differ in other diffs)

• runs of adds or removes should be grouped
  • 4.4 - PG 14 (15)
• images where only image relative path changes - don't flag as changed
  • ie, all railroad diagrams
  • 7.3.3 - pg 42
• long word-highlighted additions or removals go off edge of page
  • Addition: 7.6.2.5.1 - pg 64
  • Removal: 12.2.2 - pg 141
• need word-diffs for long code sections
  • (or line by line?)
  • 11.3 - pg 133
• after a certain point, sections flagged as substitutions that seem unchanged?
  • ie, starting in 13.3.2, pg 199

For reference, find all diff documents in this zip:
diffs_v1.0.1_to_v1.1.0-original.zip

...

[asluk]

Adding data + a possible approach for bullet 2 ("images where only image relative path changes → ie, all railroad diagrams")

While prepping a diff for the core-spec-wg v1.1.0 ballot I measured the scope of this one and found a clean way to separate real diagram changes from the path-only churn — recording it here for whenever bullet 2 gets picked up.

Dataset: --diff of core-spec-wg 2e781be (tagged v1.1.0) → cb1ac59 (post-core-spec-wg#519, an editorial-only grammar rule-name rename in a single file).

Measured ground truth: decoding the railroad SVGs from the before/after renders and comparing them at the vector-<text> level (immune to path/layout churn): 12 of 209 diagrams actually changed; the other 197 are noise. The 12 match #519's source diff exactly (renamed rules + their references).

Approach for the fix: a content-based comparison — hash/normalize the image content rather than keying on the path — is sufficient to distinguish real diagram changes from path-only churn, at least for the SVG railroad diagrams (the bulk of the figures). I have a throwaway script that does exactly this (decode embedded SVG → diff <text> → annotate the changed labels); happy to share if useful as a reference.

Related: filed #102 for a separate PDF-build fragility (unvalidated ephemeral SVG→PDF conversion) — distinct from this diff issue, but both touch how diagrams flow through the pipeline.