Count and replace from one set of instances

[forgetso]

Closes #135.

Counting measured an element's innerHTML as a string; replacing drove a TreeWalker. They disagreed, and the disagreement was visible.

Note

Based on #133, not master, because both touch searchreplace.ts. Sibling of #134 rather than stacked on it.

The measurement

Page with four occurrences — visible text, a JSON <script>, a <style> rule, an href:

	before	after
count, Replace HTML on	2	4
count, Replace HTML off	1	1
remaining after each Replace Next	4 → 3 → 2 → 1 → 0 ✗ (was 2 → 1 → 1 → 1 → 0)	4 → 3 → 2 → 1 → 0

All four were always replaced; only the count was wrong. Two presses in a row left the number unchanged, which reads as "I pressed it and nothing happened".

Why patching one side doesn't work

I tried the one-line version first: make the count read the unfiltered original so it stops missing the script and the style. That takes the invariant suite from 4 failures to 13 — the count then disagrees about hidden text and input values instead.

An innerHTML string cannot express scope. The text of a hidden element, of a script and of an input are all in it, whether or not any of them is in scope for this search. So the fix has to move off strings.

One list, three kinds

collectInstances() walks the page once and returns where the term was found:

kind	written through	in scope
text	node.data	always — covers prose, and script and style contents
attribute	setAttribute	only when replacing HTML
markup	innerHTML	a match that exists only in the tags, e.g. searching <p

count is the size of the list, Replace All walks it, Replace Next takes the first entry and replaces exactly one occurrence. The count cannot drift from what replace will do because there is one source of truth.

Two more fixes that fall out

Replacements stop destroying subtrees. Writing through the node holding the match means an ancestor's innerHTML is never reassigned, so the page keeps its listeners, selection, focus and framework DOM identity. New invariant: a match in one element does not recreate its siblings.

"Hidden content" now means something in HTML mode. The known limitation recorded in the invariant tests — that Replace HTML rewrote hidden text regardless of the option — is fixed. The test that pinned it now asserts the opposite, with a companion showing hidden text is rewritten when the option is on.

Markup matches still work

Searching <p, or <span>x</span> → <b>y</b>, is a real Replace HTML capability and two existing tests cover it. Anything in an element's innerHTML that the text and attribute instances cannot explain is charged as a surplus to the innermost element that can account for it, walking in reverse document order so a descendant is considered before its ancestor.

An element whose subtree holds anything out of scope never gets one: its innerHTML describes that hidden text or that script too, and rewriting it would drag them back into the replacement. That taint check is what stops the markup fallback reintroducing the bug it sits next to.

Application order matters for the same reason — a markup instance recreates its subtree and detaches the text nodes other instances hold, so everything else is written first and markup instances go deepest-first.

Deletions

replaceInner, nodesUnder, isIgnored, equivalentInIgnoredElements, containsAncestor, getHiddenElements. The clone-and-filter machinery existed only to keep unwanted subtrees out of an innerHTML rewrite; nothing is rewritten wholesale any more. Net −279 lines of the old path.

Checks

npm run checks passes — typecheck, lint, format, 787 tests, build. 16 new, including the invariant matrix gaining "each Replace Next takes exactly one match off the count", which now holds for every page shape × option set.

Known edge

Escaped characters. A text node's data is unescaped where innerHTML is not, so searching for a bare & or < can make the markup surplus arithmetic over- or under-count by one on an element containing escaped entities. It is clamped at zero so it cannot produce a negative count. Worth a follow-up if anyone hits it; I did not want to widen this change further.

🤖 Generated with Claude Code