tbwikiminifier module - CSS/JS minify using wikimedia/minify

[the.rampage.rado]

thirty bees integration test report

Measured on 2026-08-24 against the local thirty bees 1.8.0 shop using PHP 8.3.30. The repository remained the canonical source. This report distinguishes measurements and observations from interpretation.

CCC and lifecycle

• Measured: installing tbwikiminifier while official tbminifier was installed returned false, showed the explicit conflict error, created no hook rows, and left both module versions unchanged. The old module was never modified automatically.
• Measured: after clean old-module removal, the new module registered exactly actionMinifyCss and actionMinifyJs for shop 1. No actionMinifyHtml row, controller, table, or setting was introduced.
• Measured: direct Media::minifyCSS() and Media::packJS() probes recorded the new hooks in Hook::$executed_hooks; outputs matched Wikimedia Minify apart from the core JS semicolon trimming/join contract. Fifteen generated CCC bundles across the tested page set parsed successfully.
• Measured: disable and re-enable removed/restored the only active hook provider in separate PHP requests. CCC CSS/JS versions advanced together through 15, 16, 17, 18, 19, 20, and 21 during lifecycle tests. After the interleaved comparison, tbwikiminifier was restored as the sole active provider and tbminifier was uninstalled.
• Observation: thirty bees core calls the first CSS/JS hook response and does not chain providers, although all responding providers can run. The explicit old-module conflict prevents the known double-provider case.

Asset corpus benchmark

Every one of the 156 generated outputs (78 inputs multiplied by two engines) passed the applicable Acorn or CSSTree syntax validation. The actual-shop subset contains 33 JavaScript inputs totalling 1,178,692 bytes and 22 CSS inputs totalling 541,778 bytes.

Actual-shop aggregate	tbminifier	tbwikiminifier	New - old
JS output	1,063,778 B	1,061,031 B	-2,747 B
JS gzip	288,557 B	287,785 B	-772 B
CSS output	424,402 B	429,813 B	+5,411 B
CSS gzip	86,794 B	87,173 B	+379 B
Summed per-input median time, JS	2,123.47 ms	896.55 ms	-1,226.92 ms
Summed per-input median time, CSS	135.19 ms	9.00 ms	-126.19 ms

The summed actual-corpus medians were 2,258.664 ms for the old engine and 905.548 ms for Wikimedia (2.49x old/new ratio). This is a minifier-operation comparison, not a front-office TTFB claim. Brotli is N/A because neither a Brotli executable nor PHP extension was available.

The 2 MiB stress cases completed successfully:

2 MiB case	Engine	Output	Gzip	Median	p95	Peak memory increase
JavaScript	tbminifier	1,825,339 B	5,380 B	4,318.185 ms	4,663.861 ms	3,653,824 B
JavaScript	tbwikiminifier	1,827,104 B	6,916 B	1,429.121 ms	1,741.703 ms	1,832,024 B
CSS	tbminifier	1,592,317 B	4,708 B	913.665 ms	949.068 ms	24,349,888 B
CSS	tbwikiminifier	1,708,828 B	5,054 B	47.876 ms	57.636 ms	8,163,352 B

Real HTTP performance

The initial sequential cold runs intentionally regenerated provider-specific CCC assets. They are valid cold observations but vulnerable to time-order/environment drift.

Cold homepage metric	tbminifier	tbwikiminifier	New - old
TTFB	617.120 ms	321.106 ms	-296.014 ms
Total	622.828 ms	327.096 ms	-295.732 ms

Provider-specific home bundles from those runs were:

Asset	tbminifier raw / gzip	tbwikiminifier raw / gzip	New raw / gzip
JavaScript	316,517 / 90,378 B	315,520 / 90,271 B	-997 / -107 B
CSS	364,698 / 68,888 B	369,260 / 69,096 B	+4,562 / +208 B

The stronger warm comparison randomized 40 blocks (20 per provider) with 10 requests per block and seed 20260824. Both modules were installed once before measurement. All 400 requests used the same page, database/shop, warmed CCC bundle state, cURL handle, cookie jar, web-server pool, and OPcache pool. Only the shop activation row changed between blocks; no lifecycle method, module install/uninstall, or CCC clear ran inside the loop. The harness cannot force Apache to select the identical worker process for every request.

Interleaved warm metric	tbminifier (n=200)	tbwikiminifier (n=200)	New - old
TTFB median	107.681 ms	106.630 ms	-1.051 ms
TTFB mean	108.977 ms	107.283 ms	-1.694 ms
TTFB p95	123.013 ms	120.897 ms	-2.116 ms
Total median	113.785 ms	112.257 ms	-1.528 ms
Total p95	128.607 ms	127.207 ms	-1.400 ms

• Measured: both engines returned HTTP 200 for all 200 samples; response size was constant within each engine (18,361 and 18,363 bytes respectively); no transport retry was needed.
• Measured: before/after CCC versions remained CSS 21 and JS 21, with the same five bundle paths, byte sizes, and SHA-256 hashes. Every transition snapshot contained exactly the requested single provider.
• Inference: warmed front-office overhead is effectively equivalent in this environment. The small 1-2 ms Wikimedia advantage should not be generalized as a guaranteed production improvement.

Browser, network, and visual validation

• Measured: home, category, product, search, CMS, cart, and checkout pages completed under both engines. Candidate product option selection, quantity change, add-to-cart, carrier change, and payment-option selection worked; no order was submitted.
• Measured: all candidate CCC JavaScript/CSS responses loaded successfully and parsed. No browser or PHP error unique to tbwikiminifier appeared.
• Observed in both baselines: the theme's external lite-youtube.min.js is an ES module loaded as a classic script (Unexpected token 'export'), and the theme bundle calls missing $.debounce. The latter prevented the mobile menu and absent search UI in both runs. These pre-existing theme defects are outside the module and were not hidden.
• Measured visual stable regions: mobile home was pixel-identical; product was pixel-identical; category changed 0.0607% of stable-region pixels with mean RGB delta 0.0357. Desktop home used different effective responsive breakpoints and cart contents were intentionally changed, so those pairs are retained but marked incomparable. The candidate product full-page image has a browser fixed-header stitching artifact below y=720, which is excluded from the stable region.

Security and dependency review

• Measured: composer audit --locked and the production-only composer audit --locked --no-dev reported no advisories. The complete development graph reports abandoned marcusschwarz/lesserphp; it is transitive development-only benchmark code and is absent from the release. Production dependencies are Wikimedia Minify 2.11.0 and pear/net_url2.
• Measured: the repository security scan found zero reportable findings. Production contains no public controller, upload, SQL, shell command, eval, deserialization, remote include, outbound request, telemetry, source logging, or source persistence.
• Observation: developer benchmark/capture utilities are CLI-only and excluded from the release ZIP.

Limitations and evidence retention

Wikimedia's deliberate parser trade-offs remain: CSS source-looking comments inside quoted strings and JavaScript multibyte line separators in certain comment contexts can change content. These have regression documentation and are correctness concerns, not remotely exploitable module attack paths. Invalid CSS that does not make Wikimedia throw cannot always trigger fallback; valid CCC inputs remain a prerequisite.

All JSON, CSV, Markdown, and screenshot results under build/reports/ are retained. Raw non-customer test-shop capture inputs under ignored build/captured/ are also retained for reproducibility and are excluded from the release package.


--------------------------

I would invite everybody to test this new module.

Simply uninstall the bundled minifier module and install this one. Test for visual bugs. The cold combined js and css files should take 2x less time to generate. Which I think is a huge win for FTTB for customers coming from SERPs directly onto your product/category pages. 3 times less memory usage is a bonus. Yes, our total combined css+js package becomes 3 kb larger but I doubt this is an issue with modern 5G and optical networks if our page is around 1 mb in total.

tbwikiminifier-1.0.0.zip