[andreybolonin]

Reading multiple Cookie fields currently joins them into a Go string, removes NUL bytes into another string when necessary, and copies the result into a C string for PHP.

Compute the filtered length and copy the fields and ; separators directly into the final NUL-terminated C buffer. Preserve empty-field handling, nil versus a non-nil empty result after NUL removal, field order, and the existing joined-length overflow checks. The buffer still uses malloc/free and is released by frankenphp_free_request_context().

Add cookie parsing regression cases for classic and worker mode: absent and empty fields, NUL-only fields, NUL within values, separators around empty fields, duplicates, and long values.

Initial standalone prototype measurements (Go 1.27.3, Apple M2, darwin/arm64, one 150ms sample per case, including freeing the result):

Cookie input	Before	After	Time reduction	Go B/op	Go allocs/op
Two fields with 256-byte values	143.5 ns	103.7 ns	27.7%	576 → 0	1 → 0
Two short fields	89.08 ns	77.85 ns	12.6%	32 → 0	1 → 0
One short field	69.32 ns	67.98 ns	1.9%	0 → 0	0 → 0
Six short fields	112.4 ns	112.4 ns	0%	32 → 0	1 → 0

These exploratory results describe the prototype before the joined-length overflow guards were restored; the final revision has not been benchmarked. Go benchmem excludes C allocations. Both implementations request the same final C allocation size, and this change removes intermediate Go strings. Whole-request latency and process RSS were not measured.

Validation: gofmt and git diff --check passed. The exact final CGO helper passed edge cases and -race in an isolated harness. Differential fuzzing against the previous Join/ReplaceAll behavior passed 580,435 executions; go vet passed. The added PHP integration cases and full suite require CI: the local environment lacks the required PHP ZTS/shared-library and watcher setup. Benchmarks were not rerun for this PR.