Include the test helper once in every test that loads it - #309

tests/php/test_helper.php declares class TestCase at file level. A persistent PHP worker keeps its class table between requests, so a test that pulls the helper in with a bare require fatals on the second request one worker serves:

Fatal error: Cannot redeclare class TestCase (previously declared in .../test_helper.php:5)

The runner sees a non-JSON body and reports HTTP 500 against whichever test drew that worker, so the failure names the wrong culprit and moves from run to run.

This replaces require with require_once in the 234 places that load the helper.

Not a fix for a live failure

Nothing currently running in a worker-mode profile still used the bare form — all 142 suite entries of the seven profiles that set WORKER_FILE (worker, hooks, hooksdb, fibers, breaker, revolt, revoltclassic) were already on require_once. What this closes is the trap: 234 files carried a form that would fatal the moment one of them is moved into a worker suite, which happens routinely, and would misreport when it did.

Three of the 234 live outside tests/php/ (fixtures/framework/index.php, fixtures/php_deny/public/index.php, fixtures/php_deny/public/_security/denied.php). They load the same helper by a different path depth, which is why a grep scoped to tests/php/ missed them.

Deliberate bare includes are untouched — in particular tests/php/breaker/test_breaker_shutdown_dtor_fatal.php, whose bare require of breaker_redeclare.php exists to trigger exactly this fatal.

Verification

Both sides observed on the hooks profile (PHP_WORKERS: "1" plus WORKER_FILE, so consecutive requests provably land on one worker). Test files are volume-mounted, so neither side needed an image build.

Before, three requests to one swept test:

#1  200  {"test":"gateway_interface","pass":true,...}
#2  500  Fatal error: Cannot redeclare class TestCase ... test_helper.php:5
#3  500  Fatal error: Cannot redeclare class TestCase ... test_helper.php:5

After, five requests, the last two to a different swept file on the same worker — so this also covers the converse, that a class declared by the first test stays visible to a later one rather than merely not being re-declared:

#1..#5  200  valid JSON, all passing

Full suite before and after, 30 profiles: 635 passed / 0 failed / 1 error (636) both times, with byte-identical counts in every profile. The single error is shared/test_channel_async_usevar_stress, a known flake unrelated to this change. The five profiles that consume the fixture files were re-run separately after those three were swept and match their baselines (7 / 3 / 1 / 1 / 1, all passing).

Host checks: cargo fmt -- --check clean, cargo clippy --no-default-features -- -D warnings exit 0, cargo test --no-default-features 955+ passed / 0 failed. No Rust source is touched.

Follow-ups filed separately

• TestCase::__construct installs set_error_handler and set_exception_handler on every request and removes neither. Under a persistent worker both stacks grow one entry per request — measured at 2, 50 and 500 requests, giving depth 2, 50 and 500, roughly 1.7 KB per request held for the worker's lifetime. register_shutdown_function does not accumulate: five registering requests produced five shutdown runs, not fifteen. This predates the sweep and is untouched by it.
• Six tests under tests/php/worker/ justify their echo style with a comment saying the helper is pulled in by a bare require. That stopped being true earlier and is now false everywhere, so it reads as a blocker that no longer exists.