Version: 1.0
Date: 2026-10-01
Status: TFP-141/142/144 done and verified; TFP-143 closed; TFP-145 done (0504f92, 4d35b28) and verified; TFP-152 done (34622b0, b15714b) and verified; TFP-151 and TFP-159 to TFP-161 done and verified; TFP-162 (cefa24d), TFP-163 (be688d3, f4ed183) and TFP-164 (837de4f) done and verified 2026-10-07. Every ticket here is closed (header updated 2026-10-19).
Running TestForge three times against a real WordPress site
(https://rizonetech.com/, read-only, build 899bdd9) surfaced four issues
in TestForge itself, separate from the site’s own defects. Evidence is in
benchmarks/2026-10-01-rizonetech/ (run ids: Standard 5ce88c9c, passive
security 4134252c, accessibility a3eb621d).
These are TestForge template/consistency faults, not rizonetech.com’s problems. They are logged here rather than fixed in place because they were found in an exploratory run, not while building a ticket; the standing execution policy’s “fix defects in-ticket” applies once each is picked up.
TFP-140 was the last issued ticket.
href="#"bug, generate, functionallocator.count: Unexpected token "#" while parsing css selector "#". It
reads a link’s href="#" and passes # straight to a CSS locator, which
is invalid. href="#" placeholder links are common, so this fails against
ordinary sites and reports a tool crash as a site failure.href="#" and href="" as “no
in-page target to resolve”, not as a selector.href="#" link and an empty-fragment link.TEST_PLAN.md case.href="#" link, the check completes and reports the
link as having no in-page target, with no thrown error.Moved word for word from docs/test-plan/GEN.md by TFP-253; the case’s row there links here.
Test: The anchor-target check never crashes on href=”#” (TFP-141; unit: src/lib/generate/mock/link-audits/anchor-target.test.ts (real browser, through the Playwright CLI); fixture: /anchors)
Expected result: On rizonetech.com (run 5ce88c9c, 2026-10-01) the check “anchor links point to existing in-page targets” crashed on 5 of 5 pages: it passed each link’s href to page.locator, so href="#" became the CSS selector # (“Unexpected token "#" while parsing css selector”) - a tool crash reported as a site failure. The target is now looked up in the page by id, or by name, with the fragment URL-decoded; no selector is built from an href. An empty fragment (#) is a named finding - ‘link “Menu” has href=”#”, which points at no in-page target (a placeholder link)’ - and not a failure; #top passes, since browsers scroll to the top for it with no such element. A real missing target still fails, as “Anchor target missing for #
bug, generate, functionalTEST_PLAN.md case covering the timeout-reason path.Moved word for word from docs/test-plan/GEN.md by TFP-253; the case’s row there links here.
Test: The hover-state and HTTPS/SSL link checks bound their own waits and name what they could not check (TFP-142; unit: src/lib/generate/mock/link-audits/link-check-timeouts.test.ts (real browser, through the Playwright CLI))
Expected result: On rizonetech.com (run 5ce88c9c, 2026-10-01) the hover check hit the 30 s test timeout on 5 of 5 pages - hover() retried forever on a skip link parked off-screen (“outside of the viewport”) - and the HTTPS/SSL check on 3, requesting every secure link in turn (59 on one page, many repeated) at up to 15 s each. Hover: each link gets 2 s and the check 15 s; a link that can’t be hovered is named with the reason - “links that could not be hovered, so their cursor was not checked: #content (outside of the viewport)” - and links left when the budget ends are a finding (“N of M links not hovered: the check stops after 15 s”). A hovered link without a pointer cursor still fails. HTTPS/SSL: each distinct URL is requested once, for up to 5 s, and the check stops after 20 s; a URL that can’t be reached is named with its reason (“no answer within 5 s”, “getaddrinfo ENOTFOUND …”) - reachability is link health’s job - and the ones left when the budget ends are named. A link marked secure that is not https, or a request that ends on http, still fails. Proven in a real browser: an off-screen skip link passes in seconds with the named finding (timed out at 30 s on the previous code); a server that accepts and never answers, listed twice, plus an unresolvable host passes in under 15 s with one finding naming both and their reasons, and the stalled URL is requested once (failed on the previous code). Live, rizonetech.com, one worker: all 10 checks pass - hover 20 to 22 s (15 s budget plus the page load), HTTPS/SSL 4.7 to 12.5 s; cnvrg.io is named as ENOTFOUND. Output measured: check:template-unchanged gives 164 of 166 identical; the 2 that differ (page-data, no areas, fast and deep) differ only in the hover test. On the rizonetech.com page data (fast and deep) only the 5 hover and 5 HTTPS/SSL tests change.
enhancement, generate, accessibilitygetByText) and the Comprehensive
axe scan’s link-name rule report the same missing-name links, and the
getByText version is the noisier of the two. On a Comprehensive run the
two overlap.TEST_PLAN.md case.image-alt 1.1.1 and link-name 2.4.4 on the same image link) is
two distinct WCAG failures; merging them would hide a real one, so they are
left separate. The real duplication the ticket gestured at lives in
Standard runs, which this ticket’s criterion never covered; it moves to
TFP-145.bug, generate, securityTEST_PLAN.md case that
asserts the two paths agree on the same fixture cookies.Moved word for word from docs/test-plan/GEN.md by TFP-253; the case’s row there links here.
Test: The Standard and passive-security checks flag the same cookies for missing HttpOnly/Secure (TFP-144; unit: src/lib/generate/mock/cookie-scope.test.ts (real browser, through the Playwright CLI); fixture: security-mixed-cookies)
Expected result: On rizonetech.com (2026-10-01) the Standard security check named 6 cookies for missing HttpOnly and Comprehensive passive security none, on the same site: the Standard check took every cookie the site set, passive security only those named like a session. Rule: in scope are the site’s own cookies named like a session (SESSION_COOKIE_NAME, the rule passive security and the cookie/session check already use, Phase 34 Decision 3). All 6 rizonetech.com cookies were analytics or consent cookies (_clck, _clsk, _fbp, tk_ai, FCCDCF, FCNEC) that their own scripts must read, so HttpOnly is not theirs to have. The Standard check is renamed “session cookies set by the site carry protective flags (recorded as findings)”, names each as ‘session cookie “
bug, generate, accessibilitystandard-suite.spec.ts,
standard-run.json, content-profile.json). Two causes, both in how the
Standard accessible-name checks locate their element:
getByRole('link', { name }) and pass. Others use
getByText(text, { exact: true }).first() and fail with an empty
accessible name: TFAC-008 “Home”, TFAC-009 “Services”, TFAC-018 and
TFAC-038 “AI”. getByText resolves to the element holding the text,
most likely the <span> inside the menu <a> in WordPress menu markup.
A span has no accessible name, so a named link is reported as unnamed.
The same “AI” link passes TFAC-005 (by role) and fails TFAC-018 (by
text)./ai-search/),
a category filter and a post’s category tag (/ai/). Each check ends in
.first(), so four checks meant for four elements test one.href
/ rawHref, containerLabel) instead of .first(). Elements that are
truly identical (same name, target and container) get one check.getByText for some elements and
getByRole for others, and record it in the commit.<span>;
two same-text links to different targets; and a genuinely unnamed link
(an <img> with no alt inside an <a>).TEST_PLAN.md case.<a href="/"><span>Home</span></a> passes its
accessible-name check.npm run check:template-unchanged shows only the accessible-name
checks changed.0504f92:
check:template-unchanged afe38a1 0504f92 gives
162 of 166 identical. Split into tests, the 4 that differ change only
TFAC-002 to 005, the accessible-name tests; headers and titles are
identical. tsc is clean, and ESLint is clean on the changed files.
accessible-names.test.ts (non-browser case) and the golden-output test
pass./names run (7 passed, 2 failed), the
live rizonetech.com run (36/4 to 40/0) and the full unit suite
(423/423). The VM has no Chromium and stops runs over 3 minutes.0504f92 and passes after the fix.0504f92: only the accessible-name
tests change.4d35b28, verified 2026-10-05 in Cowork:
check:template-unchanged 0504f92 4d35b28 gives
162 of 166 identical. In the 4 that differ, only TFAC-002 to 005
change. The non-browser case and the golden-output test pass.0504f92 and passes on 4d35b28; unit suite 424/424. On the saved
rizonetech.com page data no visible link’s position changes, so the live
40/40 from 0504f92 stands for the visible checks. Only hidden copies
move, and their checks only test that the link is attached.Moved word for word from docs/test-plan/GEN.md by TFP-253; the case’s row there links here.
Test: The Standard accessible-name check tests the link it was generated for (TFP-145; unit: src/lib/generate/mock/actionable-elements/accessible-names.test.ts (introspects a real page, then runs the generated checks through the Playwright CLI); fixture: /names)
Expected result: On rizonetech.com (run 5ce88c9c, 2026-10-01) TFAC-008 “Home”, 009 “Services”, 018 and 038 “AI” failed as “has no accessible name” on links that have one. Introspection records a name only for an element with an aria-label or a <label> (its aria-label ?? label ?? text chain stops at the label, which is "" when there is none), so most links fell through buildLocatorExpression to getByText(text, { exact: true }).first(). That resolves to the element holding the text - the <span> inside a WordPress menu link, which has no name - and .first() made same-text links all test the first one. A link with a target is now found the way introspection recorded it: the <a> whose resolved href and collapsed text match, and which of those it was in page order; a link that is no longer there fails as “
bug, generate, accessibilitygetByText(text, { exact: true }).first(). When its text sits in a child
<span>, the check reads the span’s name and fails a named button. Same-text
elements still all test the first match.<div> with no role for an accessible name. A generic element has
none to give, so asking is the check’s mistake. The check must not ask an
element with no role for a name.buildLocatorExpression and the introspection name chain as they
are, unless a measured case shows that’s not enough.TEST_PLAN.md case.<button><span>Save</span></button> passes its accessible-name check.npm run check:template-unchanged shows only the accessible-name
checks changed.bf9cc36, verified 2026-10-06 (Cowork):
check:template-unchanged 91222c5 bf9cc36 gives 166 of 166 identical. No
fixed input generates a button or form name check.
signup-form check. A form with an id but no name has no role, so the
check is now Not run with its reason (T23).Moved word for word from docs/test-plan/GEN.md by TFP-253; the case’s row there links here.
Test: The accessible-name check tests the button, field or form it was generated for, and doesn’t ask an element with no role for a name (TFP-151, T23; Not run: owner decision, 2026-10-06; unit: src/lib/generate/mock/actionable-elements/control-names.test.ts (introspects a real page, then runs the generated checks through the Playwright CLI))
Expected result: Lookup (accessibility.ts): a button, select, textarea or input with no id, label, placeholder or role-and-name - until now getByText(text).first(), the element holding the text - is found among its own kind (button, input[type=button|submit|reset], [role=button] not on <a>; select; textarea; other inputs) by the text introspection recorded (computed as introspection does: an input’s value, aria-label, placeholder or name, else collapsed textContent), its visibility and its position among those, as TFP-145 finds links. A <form> without an id is the <form> itself, by TFP-160’s record (action, method, occurrence), or the nth <form> in a profile without one. Elements with an id, label, placeholder or role-and-name keep their locator. No role: a <form> with no aria-label, aria-labelledby or title (no role attribute) is generic under HTML-AAM, so its check is Not run: “
bug, generate, accessibilitylink-name rule flags the same link. Its only content is an
image with alt="" and a title.
alt
attribute “even if its value … is the empty string”, says such an
image “is mapped to the presentation role. It has no accessible name”,
and uses title only “if … there is no alt attribute”. So the link
has no accessible name.alt="" image presentational only when it has no
title, and its name code falls back to the title when alt is
empty. toHaveAccessibleName therefore finds a name.title is on the <img>, not the
<a>. A title on the link itself is a valid name (accname step 2I),
and then Playwright is right.Accessibility.getPartialAXTree), as evidence only. The verdict follows
HTML-AAM either way, decided before measuring.title of an
alt="" image, citing HTML-AAM 4.1.10 in the message. Approach
decided by Emil on 2026-10-05: the standard decides, not any one tool.title, the false pass would
stay. It would also tie generated suites to Chromium and change every
accessible-name check.TEST_PLAN.md case.docs/decisions/data/labelled-failures.md as a
template error: a false pass.<a href="/"><img alt="" title="Home"></a> fails its accessible-name
check, naming the link and HTML-AAM 4.1.10.<a href="/" title="Home"><img alt=""></a> and
<a href="/"><img alt="Home"></a> pass.npm run check:template-unchanged shows only the accessible-name
checks changed.34622b0:
check:template-unchanged cdec8e0 34622b0 gives
162 of 166 identical. Split into tests, the 4 that differ change only
TFAC-002 to 005, the accessible-name tests; headers and titles are
identical.benchmarks/2026-10-05-tfp152-logo-link/):
the title is on the <img alt="">, not the link. Playwright names the
link from it, and Chromium 147’s own accessibility tree gives it the name
"". So the standard and Chrome agree, and Playwright is the one that
differs. On rizonetech.com, / gives 39 passed and 1 failed (TFAC-007).
Unit suite 437/437. Labelled as T18, the set’s first false pass.title in the page before reading the
name. Each test loads its own page, and nothing is sent to the site.a[href]
first). The false pass on the logo dates from 0504f92, as T18 records.img[alt=""], so it misses an alt that is only whitespace, such as
alt=" ". HTML-AAM 4.1.10 treats an alt that is empty “when trimmed
of whitespace” the same way. Playwright 1.59.1 treats alt=" " as an
image and falls back to the title, so <a><img alt=" " title="Home"></a>
still passes falsely. The check must match any img whose alt is
present and empty after trimming, with a real-browser case for
alt=" ".b15714b, verified 2026-10-06 in Cowork:
check:template-unchanged 34622b0 b15714b gives 162 of 166 identical,
and only the Accessibility block changes. The alt=" " case is measured
by Code: a false pass on 34622b0, a failure now.Moved word for word from docs/test-plan/GEN.md by TFP-253; the case’s row there links here.
Test: The accessible-name check follows HTML-AAM 4.1.10 for an <img alt="">, not Playwright’s departure from it (TFP-152; owner decision, 2026-10-05: the standard decides, not any one tool; unit: src/lib/generate/mock/actionable-elements/accessible-names.test.ts (introspects a real page, then runs the generated checks through the Playwright CLI))
Expected result: HTML-AAM 4.1.10 (checked 2026-10-05) gives an image with alt="" no name and uses title only when there is no alt attribute; Playwright 1.59.1 falls back to the title. Before reading the name, the check takes the title off every img[alt=""] inside the element (the page is not reused). When that leaves no name it fails as “ no name and does not use its title ("
` fails with that message, naming the link by its target; `
` (accname step 2I) and `
` pass. On the previous code all three passed. **rizonetech.com, live:** the logo link (TFAC-007) is exactly that markup; Chromium 147's own accessibility tree gives it the name `""` (evidence only). The 40 checks for `/` from the saved page data give 39 passed and 1 failed, TFAC-007 with that message (`benchmarks/2026-10-05-tfp152-logo-link/`; labelled-failure T18). **Output measured:** `check:template-unchanged` against `cdec8e0` gives 162 of 166 identical; the 4 that differ are identical once the accessible-name tests are removed, and all 16 accessible-name tests in them change. The golden file changes only in its accessibility key. **Follow-up (2026-10-06):** "empty" is HTML-AAM's "empty when trimmed of whitespace": any `img` whose `alt` is present and only ASCII whitespace counts, not only `alt=""`. Playwright treats `alt=" "` as a real image and falls back to the title. **Proven in a real browser (boundary value analysis on the alt):** `
` passed on `34622b0` and now fails with the HTML-AAM 4.1.10 message; the other three cases are unchanged. **Output measured** against `34622b0`: 162 of 166 identical; the 4 that differ change only in their 16 accessible-name tests; the golden file only in its accessibility key. Other places where Playwright departs from the standard are not covered; once TFP-148 adds the axe scan to Standard, a disagreement shows in the report.
bug, generate, functionalhref and its presence in the page,
without requiring it to be visible. They don’t wait on the keyboard or
mouse checks, so a broken menu never hides a dead link, and links in a
broken menu are still checked.TEST_PLAN.md cases.npm run check:template-unchanged shows only these checks changed.6852538, verified 2026-10-06 (Cowork):
check:template-unchanged 882a47d 6852538 gives 162 of 166 identical.
The 4 that differ change only TFRM-001, the responsive check; no fixed
input has a hidden link. Measured by Code: the new case reproduces T19
and T20 on the old code, and the unit suite gives 445/445.
Moved word for word from docs/test-plan/GEN.md by TFP-253; the case’s row there links here.
Test: A link in a closed menu or behind a collapsed navbar’s toggle is checked as what it is (TFP-159, approach: owner decision, 2026-10-06; unit: src/lib/generate/mock/link-audits/hidden-links.test.ts (introspects the comparison fixture, then runs the generated suite through the Playwright CLI, deep mode); fixture: menus, menus-toggle-broken)
Expected result: Rule (link-audits/reveal.ts, run in the generated test): the control that reveals a hidden element is an aria-expanded element whose aria-controls (or data-bs-target, data-target, href="#id") names a hidden ancestor, failing that one just before that ancestor (Bootstrap’s dropdown), nearest hidden ancestor first; a hidden control is revealed first, up to 3 levels. Link health (a link hidden at inspection): its presence (toBeAttached, located with includeHidden) and its own address (HTTP status under 400, or its href), never its visibility; it then tries to reveal the link by click and, if it can’t, records a finding: “ is healthy by its address, but users can’t reach it: the ‘<menu>’ menu didn’t open” (or “… opened, but it stays hidden”, or “no control was found that reveals it”). Link type: a hidden link is classified by its address and presence. Keyboard: opens the menu by keyboard (focus the control, Enter); a link still hidden fails, “ can’t be reached by keyboard:
bug, generate, functional<select> matched its own options’ joined text;<form> element
itself, by what introspection recorded (its action, its fields), not by
text.<select>, and TEST_PLAN.md cases.<select> form each find
their own <form> and count its fields correctly.npm run check:template-unchanged shows only the form checks changed.91222c5, verified 2026-10-06 (Cowork):
check:template-unchanged 6852538 91222c5 gives 162 of 166 identical.
The 4 that differ change only the two [form] checks (TFFU-004 and
TFFU-008), and no test is added.
Moved word for word from docs/test-plan/GEN.md by TFP-253; the case’s row there links here.
Test: The form check finds the <form> itself and the fields introspection recorded in it (TFP-160, approach: owner decision, 2026-10-06; unit: src/lib/generate/mock/actionable-elements/form-check.test.ts (introspects the comparison fixture’s /forms, then runs the generated checks through the Playwright CLI); fixture: /forms, forms-missing-field)
Expected result: Introspection records, for each <form>, a form record: its action and method as written (“” when absent), its occurrence among the page’s earlier forms with the same action and method, and its fields (input other than hidden, select, textarea, button: tag, type, name, id) in document order. The check finds, among the page’s <form> elements, the one with that action and method at that occurrence (failing “no form on
bug, generate, seo/, the page that links to every example. Every other page
then looked unlinked and failed as an orphan.TEST_PLAN.md cases.f4bb043, verified 2026-10-06 (Cowork):
check:template-unchanged bf9cc36 f4bb043 gives 166 of 166 identical.
requestedPaths, so old profiles keep the old
check, and only new inspections get the fix.#top) no longer count as inbound
links, as the check’s title implies.Moved word for word from docs/test-plan/GEN.md by TFP-253; the case’s row there links here.
Test: The orphan-page check judges only pages that were loaded (TFP-161, T24; owner decisions, 2026-10-06: inspection records requested and failed pages, the start page is the first path requested, a 5xx is not loaded; unit: src/lib/generate/mock/link-audits/orphan-pages.test.ts (introspects the comparison fixture, then runs the generated check through the Playwright CLI); fixture: hub-down, /unreachable)
Expected result: Inspection (inspectPages, /api/introspect) adds requestedPaths (in order) and failedPages ({ path, reason }) to the profile. A page fails when it can’t be loaded (timeout, dropped connection) or answers 500-599 (“HTTP 503”); it is left out of pages. A 4xx is still a page. The check, for a profile with requestedPaths: the start page is the first requested path. Start page not loaded: Not run, “the start page / could not be loaded when the site was inspected (HTTP 502), so which pages it links to is unknown”. Otherwise each loaded page except the start needs a link from another loaded page; a page’s links to itself (#top) no longer count. With another page not loaded, an unlinked page is a finding (“/anchors has no link from another loaded page; /unreachable could not be loaded, so a link from there was not seen”), not a failure; with every page loaded it fails, “/anchors is not linked from any other inspected page”. A profile without requestedPaths (saved before TFP-161) gets the previous check, unchanged. Proven (decision table: start loaded or not; another page lost or not; page linked or not), 2026-10-06: hub-down with /, /anchors, /names (pages only the home page links to, like the-internet’s examples): Not run, naming /; the previous check on the same profile failed /names as an orphan, as T24. /, /anchors, /unreachable: passes with the finding naming /unreachable. /, /notes, /anchors: fails naming /anchors. /, /login, /notes: passes, no finding. A profile without the record gets the previous check. Output measured: check:template-unchanged against bf9cc36: 166 of 166 identical. The fixed inputs’ stored profiles have no requestedPaths, so they keep the previous check; the new path shows only in profiles inspected from now on.
bug, generate, functionalgetByRole, which only finds elements in the accessibility tree, so it
never finds the button.TEST_PLAN.md cases.check:template-unchanged shows only these checks changed.cefa24d, verified 2026-10-07 (Cowork), in the batch with TFP-177.Moved word for word from docs/test-plan/GEN.md by TFP-253; the case’s row there links here.
Test: Hidden elements in the remaining checks (TFP-162, TFP-159’s approach; unit: src/lib/generate/mock/link-audits/hidden-links.test.ts)
Expected result: Deep mode’s navigation destination check: a link hidden at inspection (a closed menu) is taken by its own address - present in the page, its href resolving to its destination - and not clicked; whether users can reach it stays the keyboard check’s question. Emitted only when the page has such a link. Form and name checks: an element hidden at inspection that has a role and a name is looked up with getByRole(..., { includeHidden: true }), since getByRole alone searches the accessibility tree, which leaves hidden elements out (0 matches against 1). Proven 2026-10-07: on the fixture’s menus state, deep mode, the destination check passes with Hand Tools and Archived in the closed Categories menu (it failed on the previous code, “Hand Tools … not found”), and the closed menu’s hidden “Clear filters” button (aria-label) is looked up with includeHidden and its checks pass.
bug, introspect, accessibility<input> gets the role textbox, so a labelled
type="search" field (role searchbox) is never found by role.Each of these changes the saved page data, so measure on the fixed inputs and on fresh inspections.
TEST_PLAN.md cases.be688d3, f4ed183, verified 2026-10-07 (Cowork), in the batch with TFP-177.Moved word for word from docs/test-plan/INT.md by TFP-253; the case’s row there links here.
Test: Introspection records controls as they are (TFP-163; unit: src/lib/introspect/controls.test.ts)
Expected result: Roles by input type (ARIA in HTML): search is a searchbox (a combobox with a list), number a spinbutton, range a slider, image/submit/button/reset a button, text-like fields with a list a combobox; text, email, tel, url and password stay textbox. Buttons: a button with no text, label or id (icon-only) is recorded, with its position among the page’s buttons, and found by it; same-text buttons are no longer merged - each records its position and which of the same-text buttons it is, and is found as the nth. Older profiles (no position) generate exactly as before. Proven 2026-10-07: each input type’s role (equivalence partitioning); on the fixture’s /controls, the labelled search field is found by role and its checks pass, the icon-only button’s name check fails (“has no accessible name”), and the two “Delete” buttons get two checks, the enabled one passing and the disabled one failing. A fresh, read-only inspection of with-bugs /#/auth/login now records the show-password toggle (R15) at position 3.
bug, generatecoverage-tests.ts (since 2026-08-28), the
pattern is written as \. inside a template string, which drops the
backslash, so the generated
test gets /.(pdf|doc|…)…$/i. The . there matches any character, so
…/getpdf counts as a download. The hidden-link path TFP-159 added
escapes it correctly.\. in the generated pattern. Add a unit case: a link
ending xpdf with no dot is not a download.\.(pdf.check:template-unchanged shows only the download check changed.Moved word for word from docs/test-plan/GEN.md by TFP-253; the case’s row there links here.
Test: The download checks’ pattern keeps its escaped dot (TFP-164; unit: src/lib/generate/mock/link-audits/download-pattern.test.ts)
Expected result: Written as \. inside a template string, the backslash was dropped, so the visible link-types check and the download-link check emitted /.(pdf|...)/, which counts .../getpdf as a download. Both now emit \.(pdf|doc|...)([?#].*)?$, as TFP-159’s hidden-link path already did. Proven 2026-10-07: equivalence partitioning on the emitted pattern - report.pdf, report.PDF, sheet.xlsx?v=2, deck.pptx#page=3 match; getpdf, xpdf, reportzip don’t (the test fails on the previous code).
32e8eae, verified.6c484de, verified.4a81655, verified.0504f92 and follow-up 4d35b28, verified 2026-10-05.34622b0
and follow-up b15714b, verified 2026-10-06.6852538, 91222c5, f4bb043), with TFP-151 (bf9cc36),
verified 2026-10-06.cefa24d; be688d3, f4ed183; 837de4f), verified 2026-10-07.837de4f, verified 2026-10-07 (Cowork), in the batch with TFP-177.