perf: memoize the browser's CDP connection

[hugo-fasone]

Context

On any suite that calls importPage(), runtime is dominated by redundant CDP
reconnections rather than by actual browser work. Measured on a project with 2 suites /
5 tests (PHP 8.3, headless Chromium in a container, PS 9.1 target): 23.67 s, of
which roughly 15 s was pure bootstrap.

Launching Chromium accounts for only 208 ms — the cost is elsewhere.

Root cause

TestsSuite::getBrowser() re-reads the socket file and issues a fresh
BrowserFactory::connectToBrowser() on every call, never reusing the instance. Two
effects compound:

Operation	Measured cost
connectToBrowser()	101 ms, constant
getPages()	205 → 506 → 406 → 660 ms (growing: every connection left open makes target enumeration heavier)
TestsSuite::getBrowser()	~102 ms
TestsSuite::getPage()	~1,250 ms (2 × getBrowser())

The multiplier lives in CommonPage::__call(), which evaluates getPage() three
times — condition, method_exists, invocation:

if (!is_null($this->getPage()) && method_exists($this->getPage(), $name)) {
    call_user_func_array([$this->getPage(), $name], $arguments);
}

And importPage() calls $pageInstance->setUserAgent(...), which does not exist on the
library's Page classes and therefore goes through that __call. Full chain:

importPage()  →  setUserAgent()  →  __call  →  3 × getPage()  →  6 × connectToBrowser
              ≈ 3 × 1,250 ms  =  3.7 s per imported page

A suite importing 3 pages thus paid ~11 s before running its first test.

Changes

1. TestsSuite::getBrowser() — reuse the connected instance

Two statics ($browserInstance, $browserInstanceSocket) plus a guard at the top of the
method. Invalidation has three conditions, so a dead browser is never handed back:

• the socket URI is gone (end of run, @unlink by ExecuteSuite);
• the URI differs from the memoized one (browser relaunched elsewhere);
• getConnection()->isConnected() is false (browser closed or crashed).

isConnected() performs no CDP round-trip — it only reads the local socket state in
wrench\Client — so the guard stays free. ExecuteSuite::getBrowser(force: false)?->close()
still retrieves the live instance to close it at the end of the run.

2. TestsSuite::getPage() — one enumeration per attempt

getBrowser() is hoisted into a variable, and getPages() is only re-issued after an
actual createPage() (the first list is otherwise already current). The ×3 retry loop and
the applyExtraHttpHeaders() semantics are unchanged.

3. CommonPage::__call() — evaluate getPage() once

Results

	Before	After
getBrowser() (already connected)	102 ms, constant	0 ms
getPage()	~1,250 ms	166 ms, then 0 ms
init() = 1 × importPage()	3,910 ms	3 ms
Smoke suite (3 tests)	10.53 s	3.37 s
Full run (2 suites, 5 tests)	23.67 s	3.60 s

Compatibility

No API change and no observable behaviour change:

• __call() still returns nothing — I deliberately did not add a return on
  call_user_func_array, since that would change the return value of every proxied call
  and is out of scope for a performance fix;
• getPage() keeps its return $pages[0] (same semantics on an empty list) and its retry
  loop;
• a dead browser now raises an error instead of being silently relaunched mid-suite. This
  is the more honest behaviour: the previous silent relaunch dropped the session and
  cookies, producing failures that were hard to interpret.

Testing

• vendor/bin/phpunit --testsuite Unit: OK (117 tests, 442 assertions) on PHP 8.3.
• Integration run against a real PS 9.1 shop: 5 tests, same results as before the patch.
• The full importPage() path (3 pages, with BO credentials) exercised separately: imports
  succeed and the resulting failure is the login assertion, as expected.

[PrestaEdit]

Huge thanks @hugo-fasone for this PR 🙏

The analysis is genuinely outstanding: root cause traced all the way down to the triple getPage() inside __call, measurements at every step (connectToBrowser 101 ms, getPages growing as stale connections accumulate, importPage blowing up to 3.9 s…), and a before/after table that speaks for itself — 23.67 s → 3.60 s on a full run, a massive win for every user of the library.

The fix is surgical and minimal:

• three-condition invalidation in getBrowser() (socket gone / socket changed / isConnected() false),
• a single getBrowser() + a single enumeration per attempt in getPage(),
• one evaluation of getPage() in __call.

No API change, and the behaviour shift is spelled out in the "Compatibility" section (a dead browser now raising an error instead of being silently relaunched is indeed the more honest behaviour).

LGTM, merging. Thanks again for the quality of the contribution 👏