Change control — Friday 7 August

What it is Last Wednesday's emergency fix stopped the site sending a shopper to a different manufacturer's part for codes it could read. It couldn't read every code — six-letter codes like BRIGGS, and any code with an underscore, fell through the same guard it had just closed. This closes it for those too: an unresolvable branded code now says the part was not found instead of guessing.

How we test it An asserted-but-unresolvable branded URL 404s instead of redirecting to a different manufacturer's part; a bare part number and a resolvable branded code are unaffected.

Evidence Measured today, same method as last week — -A crop-audit, reading the served title, not just the status code.

Scope, from Tuesday's production measurement: 1,012 URLs across 20 vendor codes — 679 exact part-number collisions, 333 truncated to a wrong maker and a wrong part like the row above. BRIGGS alone is 355.

Roll back or fail forward Fail forward. Only the fail-open branch of the redirect guard changed; 58 of 58 tests pass, four independent people tried to break it and could not. Revert the merge commit if a code that should resolve starts 404ing instead.

What it is On a phone, the parts page and every brand page loaded to an empty blue banner. The first product card sat one pixel below the visible screen — not a slow load, the layout itself pushed every part out of view until you scrolled.

How we test it At the exact screen size Google's mobile index renders at (390×844), a product card — image, price, Add to Cart — is visible without scrolling.

Evidence Same URL, same viewport, before and after:

Roll back or fail forward Roll back — layout only, nothing stored.

What it is For vendor codes the search index couldn't recognise (BRIGGS, AGCO and others), the part a shopper would actually add to cart carried a bare number instead of the branded one — the inverse of the redirect bug above, on the search side rather than the page side.

How we test it Search a number that collides across manufacturers; every card carries its own manufacturer's branded number, none falls back to a bare part number.

Evidence Confirmed live on the dev backend after deploy, not just on green CI — revision search-nest-00268-pev, latestReady == latestCreated, 100% traffic, image tag matching the merge commit exactly.

Roll back or fail forward Roll back — one deploy to the previous revision, no data changed.

What it is The catalog offers 1,000 pages of parts. Past roughly page 900 the page sometimes came back completely blank — no parts, no message, and the page still reported itself to Google as a good page. It happened when the parts list took longer than five seconds to build: the count of results still arrived, the results themselves did not, and only page one was set up to notice that and say so. Now any page at any depth that cannot be built says "busy, try again" instead of showing an empty shelf, and Google is told not to keep the blank version.

How we test it A deep catalog page must either come back with a full grid of parts or come back as a busy page — never as a normal page with zero parts on it beside a results count in the hundreds of thousands.

Evidence We loaded a deep catalog page on both sites and counted the product cards in the page that came back, then repeated the load six times against each search service. The blank state is a timing race, so it did not reproduce during this measurement — that is recorded literally below rather than claimed as proof it is gone.

Roll back or fail forward Rolling back is a single swap on the search service and restores the previous behaviour, which is the blank page. There is no front-end change to undo.

What it is The brand filter in the search sidebar counted parts that are not for sale on the site, so the number on the button promised far more than clicking it delivered — on a search for FASTENERS the Ford button said 22,466 and the results said 6,268. The stock filter had a second problem: several internal stock codes all mean "out of stock" to a customer, and each one was drawn as its own separate button, so the same option appeared several times over. Now the brand count is built from the same set of parts the results grid uses, and the stock options are combined into one button each.

How we test it For one search, the number on a brand button must equal the number of results you get after clicking it, and no stock option may appear more than once.

Evidence We asked both search services for the filter panel and the result count for the same search, FASTENERS, and read the numbers back literally.

Roll back or fail forward Rolling back is a single swap on the search service and brings the inflated brand numbers back. Two other filter numbers — the category buttons and the overall total — are still overstated because they come from the database side, which is a separate piece of work already filed.

What it is When a shopper sorts search results by best sellers, page two can repeat a part that was already on page one, and a different part is never shown at all. The cause was that the set of parts we picked from grew larger the deeper the shopper paged, so each page was really a different ordering rather than a continuation. The change makes that set depend only on what was typed, so every page is drawn from the same ordering. Nothing changes for a customer yet: this is the code half only. The matching database change is Alex's to apply, and until Alex applies it the site behaves exactly as it does today.

How we test it Ask for page one and page two of the same best-sellers search and check that no part number appears on both.

Evidence We asked both the live site's search service and the candidate one for pages one and two of "oil filter" sorted by best sellers, and compared the part numbers returned; today they are identical, which is the proof that nothing has changed for shoppers yet.

Roll back or fail forward There is nothing to roll back today, because no shopper-visible behaviour has changed. Once Alex applies the database change, the previous version of that one database routine can be put back on its own, without touching anything else on the site.

What it is A sitewide audit on Monday found the breadcrumb trail (the "Home > Parts > …" strip, and the matching data Google reads) was hand-built per page in most places, so the visible trail and what Google sees could disagree. Four specific defects it found are fixed: a diagrams page that told Google it had a parent it never linked to; a mid-trail entry missing on the not-found page; two rendering engines for the same breadcrumb; and the "Parts" page named two different things on the same page. One policy change rides with it: trails now always start with "Home". That reverses a narrower decision made in June (issue #1201) that had removed Home from six page types on purpose — worth saying plainly rather than letting whoever reads the old decision next re-revert this by mistake.

How we test it A new automated check reads every page's own visible trail and demands the machine-readable version say the same thing, computed fresh rather than compared against a fixed list — so a new page with no trail at all fails the build instead of shipping unchecked.

Evidence Three of five defects shown below, each re-measured today:

The guard itself is green on this branch (CI job pr-localhost-crawl).

Roll back or fail forward Fail forward preferred — reverting brings back a real, if smaller, defect (the borrowed-trail bug is a genuine Google violation); a clean revert is available if the Home-first change needs to be undone on its own.

What it is Any made-up word after /diagrams/type/ on a brand page loaded as a normal page — an unbounded number of fake pages Google could have tried to index.

How we test it A type that does not exist answers "not found", not "found".

Evidence

Roll back or fail forward Roll back — one route file, nothing stored.

What it is A brand's banner was as tall as its intro paragraph happened to be, so it varied by up to 23 pixels brand to brand for no reason a shopper would notice a reason for, and the two components that render it disagreed on heading size and spacing besides.

How we test it Two brands using the same layout render the same banner height regardless of blurb length.

Evidence Banner height measured today on both sites at phone width (390 × 844), read off the rendered page:

Same two brands, same phone size, before (still varies) and after (fix on dev):

Roll back or fail forward Roll back — styling only, nothing stored.

What it is On the brands page the first row of brand pictures was marked as "can wait", even though it is the first thing a visitor sees, so the browser held those pictures back before starting to fetch them. They are now fetched immediately, and the browser is told about them at the very start of the page. Separately, the Clinton Tractor logo that appears on every page had a web address that changed with every release, so a returning visitor downloaded it again each time instead of reusing the copy already on their machine; the address is now fixed. Both changes are about how quickly the page appears, not about what it says.

How we test it Fetch the brands page as a visitor and count how many of the first-screen pictures are still marked "can wait", and read the logo's address to confirm it carries no release stamp.

Evidence We requested https://crop.clintontractor.net/brands and https://crop-dev.app/brands on 2026-08-06 with our audit identifier and counted the markings in the page we were served.

Roll back or fail forward Both parts are small and independent, so either can be put back on its own without affecting the rest of the page. The wording, prices and links on the brands page are untouched, so a reversal would only make the first pictures slower again.

What it is The site has a step that shrinks a picture to the size the visitor's screen actually needs. For the last 35 days that step quietly skipped two whole groups of pictures — the ones stored with the site itself, and the ones in our own four Google storage buckets, which is most of the brand and equipment photography. Those were sent at full size to every visitor, on every device. They now go through the shrinking step like everything else.

How we test it On the brands page, our own pictures are requested through the resizing address instead of straight from the storage bucket.

Evidence Measured today with -A crop-audit on /brands, counting distinct picture addresses in the page the server sends:

The change's own accountability record measured the effect on the same page at 5.03 MB and 18.5 seconds to the main picture on prod, against 1.61 MB and 9.8 seconds on dev. That is a page-speed measurement, not a status code, so it is quoted from the change rather than re-measured here.

Roll back or fail forward Roll back — it is one routing rule, and reversing it only makes pictures large again. Nothing is stored and no address a customer can bookmark changes.

What it is We used to have two general "all brands" diagram directories sitting outside the brand and equipment pages. On the live site those addresses still open, and some of them quietly forward somewhere else. On the candidate site they are retired: anyone opening one is told plainly that the page is gone, and nothing forwards anywhere. A saved or shared link to any of those addresses now says the page is gone and does not forward, so the visitor has to start again from the brand or equipment page. Diagrams themselves are not going away: they stay on the brand page at /parts/brand/{brand}/diagrams and on the equipment page at /equipment/{machine}/parts-diagrams.

How we test it Every retired address must answer "gone" with no forwarding address, and the two surviving diagram addresses must still open normally.

Evidence We asked each address directly on both sites today, one request per address, and wrote down exactly what came back.

Roll back or fail forward Undoing this means putting the directories and their forwarding rules back, which is a larger change than switching a setting. If a retired address turns out to matter, the faster answer is to add a single forwarding rule for that one address to the brand or equipment page that replaced it.

What it is

The fitment list on every part page is now a real table, with a caption and proper column and row headings. Frequently-asked questions are written as questions and answers. The supersession chain — the older part numbers that lead to the current one — is written as an ordered sequence rather than three boxes in a row. Every section of every page now has a name, and the heading levels run in order without skipping. Five separate changes, merged over two days. Tracked as front#3426, seven commits.

How we test it

The page is fetched as Google fetches it and the labelled pieces are counted, before and after, on the same addresses with the same instrument. Alongside that, every page that changed is photographed at desktop and phone width, before and after, and the two are compared — because the whole point is that nothing a customer sees should move.

The evidence

Part page /parts/CT-NHL-219758: 0 fitment tables before, 1 after; 12 unnamed sections before, 0 after. Full counts in the next row, photographs in the one after it.

Roll back or fail forward

Roll back, in one button. The storefront alone changed; no data was touched, so going back restores exactly the previous page and nothing else.

The same part page, fetched from the live store and from the staging copy within the same minute. Every number below is a count of labelled pieces the page hands to a search engine.

The six headings are worth one sentence, because it is the number that shows the table is real rather than merely present: two of them name the columns and four name the rows, against four ordinary cells — four machines, one fact each. The two remaining columns, the years a machine was built and its circuit, are left out rather than shown blank, because we do not have that data. That gap is written up separately and it is Alex's.

The same pattern holds on the other pages that changed. The store's opening hours and the figures on the About page are now written as terms and their values rather than as text; the catalogue results are a list; the diagram index is a list. Across every page checked, eight headings that skipped a level became none.

Each pair is the same page: the live store on the left, the same page after the change on the right. Full length, not a crop.

Live store nowAfter the release

Live store nowAfter the release

Live store nowAfter the release

Live store nowAfter the release

Three more pages were photographed the same way, at both widths. The home page (after) and the machine page (after) are unchanged to the pixel. The About page (after) is 43 pixels taller, and that is the breadcrumb trail described further down this page, not this change — the figures band above it is identical.

The machine page is the one that nearly went unmeasured. Its specification rows are now written as terms and values rather than as text — the live store has none of those on that page, the candidate has five — and it was not in the first round of photographs at all. It was found by re-reading the list of what changed against the list of what was checked, which is the only reason it is here.

One page dropped out of this comparison entirely: the general diagram directory at /parts-diagrams. It was photographed at 200 OK on both copies in the afternoon and was retired to 410 Gone on the candidate an hour later, by the change described further down this page. There is no after to compare it against, so it is not shown.

The fitment search box

The box that filters the fitment table by machine reaches into the page by name and had no automated test at all. Driven by hand on both copies today, on /parts/CT-NHL-219758:

Fitment box	prod today	crop-dev.app
Models listed, no filter	4 models	4 models
Type 82C	1 model, "Confirmed fit — 1 matching model below"	1 model, "Confirmed fit — 1 matching model below"
Type zzzz	0 models, "No match in this list. Check your serial number or ask our parts desk."	0 models, "No match in this list. Check your serial number or ask our parts desk."

The phone layout of the supersession chain

Below a certain width the chain of older part numbers has to stack downwards and its arrow has to turn to match. That is the one piece of the rewrite that could have broken silently. It is the second pair of photographs above.

What was not checked by an independent person

The author of the change is also the person who verified it. For a change with no visible effect that is a smaller risk than usual, but it is worth stating rather than leaving implied.

Reproduce it: bun run scripts/semantic-census.ts — fetches both copies of the site as a search engine would and prints the table above for every page. Has to run before Friday's release; after that both copies are the same and there is no longer a "before" to compare against.

What it is Until now the page the server sent contained the logo and an empty bar where the menu should be, and the menu only filled in once the browser had finished running the page. Anyone who could not run that code — Google's crawler on a bad day, a slow phone, a visitor with scripts off — saw a header with nothing in it and no way to reach the brand pages. The server now sends the menu itself, and the search box works as a plain form even with nothing running. Weekend work.

How we test it We ask the server for a page as a robot would and count the menu links it actually contains, without letting any code run.

Evidence Fetched the same page from both sites as the crawler, and counted the links inside the header of the raw text the server returned.

Roll back or fail forward It is a self-contained change to the header, so it can be lifted back out on its own and the site returns to today's behaviour. Nothing else depends on it.

What it is Every page tells a search engine and a screen reader which part of it is the main content. Fifteen places in the site were each declaring their own, so on a number of pages two blocks both claimed to be the main one and nothing could tell which to trust. Screen reader users landed in an ambiguous page and Google had to guess what the page was about. Now there is exactly one on each page.

How we test it We count how many times a page claims a main content area in the text the server sends; the answer must be one.

Evidence Fetched each page from both sites as the crawler and counted the main-content markers literally.

Roll back or fail forward Nothing a customer can see changes, so removing it is safe at any time. Two automatic checks now fail if a second main area ever comes back.

What it is The page listing every machine model lived at /equipment/all and was a 4.7 MB page that Google crawled once in June, never indexed, and that earned no clicks. It now lives at /equipment/models, is half the size, is grouped by brand instead of by a letter axis that put 96% of models under two letters, and opens with its sections expanded rather than all collapsed. Every one of the model links survived the move. The old address is gone rather than redirected, because it had no traffic and nothing links to it.

How we test it The new address answers, the old one is gone, and the new page still carries the same number of model links.

Evidence Fetched both addresses on both sites as the crawler, and counted the distinct model links and the page size in the raw text returned.

Roll back or fail forward Rolling back restores the old address and the old page whole. If the smaller page still does not get indexed, the next step is splitting it into one page per brand, which the new address can host and the old one could not.

What it is A customer with a DIS charge account can check out against that account instead of a card. The storefront half of this landed too, so the checkout review page can offer paying on the Clinton account, but it stays completely hidden until the feature is switched on, which is John's call and has not happened anywhere. With the switch off both copies of the site read identically, which is the measurement below.

How we test it A mapped DIS customer completes checkout with the account option; an unmapped customer never sees it; the route 404s while the flag is off.

Evidence Not yet walked by anyone other than the author. This is a new customer-facing payment path — it needs Denis's QA before Thursday's call, not a claim from this page.

Roll back or fail forward To be stated once QA'd. Do not ship this row without it.

Nothing a customer or a search engine can see 28

Change	What it is	Before	After	Built by
front#3338	Tidy-up of the shared-cart code after review, with no change to what a shopper sees.	same item-count sum copied in three places	one shared helper, 20 fewer lines	Vova
front#3365	Records and accepts two security warnings that have no fix available yet, so the build check stops blocking everyone.	security check failed on every branch	security check passes, two warnings recorded	Vova
front#3362	Corrects which build machines are allowed to save and reuse their working files.	machines picked by operating system, one wrong	machines picked by name, all correct	Vova
front#3375	The deepest automated test run now runs only the tests it owns, split four ways to finish in time.	lane ran 106 tests, never once succeeded	lane runs 51 tests, split four ways	Vova
front#3386	Code-style warnings now stop a change from merging instead of being ignored.	style warnings passed the check	style warnings fail the check	Vova
front#3387	Copies the last four leftover sets of records from the old database into the new one.	four record sets missing from the new database	36 records copied across	Vova
front#3319	A pickup-promise test was reading the store-hours widget, so it passed or failed by the time of day.	test result changed with store opening hours	test reads only the page content	Denis
front#3320	Gives one test step the same waiting time as the steps after it, so it stops failing at random.	step allowed 1000 milliseconds, failed randomly	step allowed 1500 milliseconds, five green runs	Denis
front#3377	Updates three outdated test expectations on the brand equipment-type page; no customer defect was involved.	3 tests failing, 9 failures a night	3 passed, 2 skipped	Denis
front#3357	Adds a test proving an order above the allowed quantity never reaches payment.	no test covered the order quantity limit	one test proves over-limit orders are refused	Denis
front#3355	A nightly check that finds broken brand pages can now report them, and prints every failure rather than the first ten.	15 failures found, 10 printed, nobody told	all failures printed, a ticket is filed	Denis
front#3382	Updates diagram-page tests to match today's product and leaves one genuine defect visible on purpose.	8 passed, 5 failed	12 passed, 1 failed	Denis
front#3391	A test file now runs in one place so its safety check stops raising a false alarm.	test failed depending on how work was split	test runs in order, check still intact	Vova
front#3383	Fixes five stale test expectations on the model-variants page and closes a mismatch that no live page could reach.	3 passed, 5 failed on the variants page	8 passed	Denis
front#3315	The staging test run now waits long enough for its own build to get its turn in the queue.	gave up after 15 minutes, ran no tests	waits up to 40 minutes	Vova
front#3393	A search test now checks for at least the expected results instead of an exact count that the live catalog keeps changing.	test demanded exactly 2 results, catalog returned 3	test demands at least 2	Vova
front#3415	Corrects a note that named the wrong service for a shared setting.	note named the retired search service	note names the live search service	Vova
front#3406	Builds the customer side of paying by charging a Clinton account, switched off everywhere for now.	no charge-to-account option existed	option built, switched off, page unchanged	Vova
front#3398	Every change must now name which assistant was used and which person is answerable for it.	no responsible name required on a change	change blocked until both names are filled	John
front#3425	Adds the instruction so assistant-opened changes fill in the newly required names.	assistant changes opened without the required names	the written flow spells the two lines out	Vova
front#3440	The nightly broken-link sweep now sees genuine not-found pages and no longer throws away filtered addresses.	sweep saw 200 on real not-found pages	sweep sees 404 and keeps filtered addresses	Vova
front#3454	Clears the last eight warnings about how files reference each other.	8 warnings	0 warnings	Vova
front#3457	Every code change must now state what it was, what it became, how to check it, and show the proof.	evidence optional unless bound for production	evidence required on every code change	Vova
front#3458	The trail checker now covers every page template automatically and runs on every change instead of nowhere.	checker covered 6 listed pages, ran nowhere	covers 37 templates, runs on every change	Vova
front#2994 no pull request, pushed straight to the shared branch	The morning sitemap check now compares the published address count against the live backend count instead of two fixed numbers that could no longer be reached.	check failed every morning, 16 repeat reports	2,694 against 2,694, zero drift	Vova
front#2992 no pull request, pushed straight to the shared branch	Deletes a test that could never fail and stops other tests calling the retired search backend when no address is set.	retired backend took about 20 requests each morning	no scheduled run calls it	Vova
front#2992 no pull request, pushed straight to the shared branch	Removes the retired search backend from the tracing list and two helper scripts.	three places still named the retired backend	none do	Vova
front#3420 no pull request, pushed straight to the shared branch	Corrects a note in the brand tile code that stated the opposite of what was measured about early image loading.	note said early loading adds no preload	note says it does add one	Vova

Nothing a customer or a search engine can see 4

Change	What it is	Before	After	Built by
search#857	A shared security key had a weaker length requirement on the search side than on the website side, so a key that looked accepted could still stop the website from starting; both sides now demand the same length.	Short key accepted by search, crashed the website	Short key rejected by both, same minimum length	Vova
search#854	Sorting search results by best sellers repeated some parts across pages and hid others; the correction is written down but only takes effect once Alex applies the matching database change, which has not happened.	No change until Alex applies the database change	No change until Alex applies the database change	Vova
search#853	The automatic checks that guard every release were failing on out-of-date expectations rather than real problems, and the clean-up step meant to remove a rejected build had never once worked, leaving old builds publicly reachable.	Release checks stuck; rejected builds stayed publicly reachable	Checks pass; a rejected build is now deleted	Vova
search#849	The code-quality check was reporting warnings but still passing, so nine of them sat unnoticed, including one that made a test pass when it should not have; the warnings are cleared and the check now stops on any new one.	Nine warnings present, quality check passed anyway	Zero warnings, check now fails on any warning	Vova